Hey Everyone,

Spotted in a the wild, yes, a purple Llama.

Somewhere in-between Google Gemini and OpenAI drama news, Meta keeps releasing these interesting AI developments and tools.

👾 What is Purple Llama Anyways?🦙

Purple Llama is an umbrella project that over time will bring together tools and evaluations to help the community build responsibly with open generative AI models. The initial release will include tools and evaluations for cybersecurity and input/output safeguards but we plan to contribute more in the near future.

Components within the Purple Llama project will be licensed permissively enabling both research and commercial usage. We believe this is a major step towards enabling community collaboration and standardizing the development and usage of trust and safety tools for generative AI development.

Why purple?

Borrowing a concept from the cybersecurity world, we believe that to truly mitigate the challenges which generative AI presents, we need to take both attack (red team) and defensive (blue team) postures. Purple teaming, composed of both red and blue team responsibilities, is a collaborative approach to evaluating and mitigating potential risks and the same ethos applies to generative AI and hence our investment in Purple Llama will be comprehensive.

Read the Blog Official

Subscribe now

Read more

Hey Everyone,

The idea that Neeva, Perplexity, Bing AI were going to challenge Google was just nuts. While Neeva was acquired by Snowflake back in May, 2023, and Bing AI sort of crashed and burned in its attempt to take marketshare from Google, Perplexity either needs substantial improvements, more funding and or to have an exit of its own.

Suffice to say that I’m not a huge believer. Its CEO likes to ramble on LinkedIn, and I wish I could say I felt more substance in the future of the product. But perhaps I’m wrong?

Recently, Generative AI startup Perplexity AI has introduced two new large language models (LLMs) that offer real-time access to the internet and information online. Perplexity's new PPLX models leverage the internet for real-time data.

Listen I like You.com and Duckduckgo, but Perplexity isn’t among my favorites.

The pplx-7b-online and pplx-70b-online models are publicly accessible via the Perplexity API and Labs web interface. Now they have partnered with Amazon.

They now leverage Llama-2 and Mistral and have found ways to optimize them. Perplexity’s online LLMs can tap into the latest information from the internet to generate responses, making them uniquely capable of answering queries that depend on recent events or data.

Perplexity is likely to get more funding soon to see if they can really achieve a Generative A.I. version of search that’s viable and competitive with Bing, Google and others.

Read the Blog (Official)

The startup is itself founded by former Google AI researchers Andy Konwinski, Aravind Srinivas, Denis Yarats, and Johnny Ho. Best case scenario for me is someone like Databricks, Amazon or Salesforce acquires them.

Their last funding round was in April, 2023 so they are very due for more funding, likely around a $500 million valuation, though this could end up being higher.

An “Answer Engine” not a “Search Engine”

Given how Google Gemini is likely to re-frame search, I’m not sure Perplexity’s model is that viable but in a gist, instead of typing keywords and sorting through a list of links, users can pose their questions directly in a conversational way and receive concise, accurate answers backed up by a list of sources.

I have not personally found the citations to be very high quality and directly related to my specific question though, so I’m not sure how viable this product is going to be.

Nor am I clear on how they would make money.

Still the way they are using RAG is interesting. Accessible through the pplx-api and Perplexity AI’s LLM playground, these models represent a novel approach in the realm of AI-driven information retrieval. The company is trying to innovate and that is important and there simply aren’t and haven’t been many search companies willing to do so. Apple might be able to undertake it seriously one day, perhaps as soon as 2025, that is make a viable alternative to Google. But I’m not holding my breath with hope.

Get started with pplx-api here.

Perplexity’s proprietary search infrastructure spanning millions of web pages. The system extracts text and data to augment the models with the latest information for time-sensitive queries.

I hope I am wrong, I do hope Perplexity can build incredible new Generative AI based queries and retrieval systems.

Origin Story

The new models continue Perplexity’s ongoing expansion of its portfolio, including its mobile app and browser extensions.

Familiar evangelists are hyping it up. Generative AI search bots offer a fundamentally different approach to the task of finding content on the internet. The ones that can take videos information and make sense of them have a lot of potential I think. There’s also a lot of untapped data in podcasts and YouTube videos that aren’t very well retrieved around niche queries.

Google’s search experience has gotten noticeably weaker in recent years. ChatGPT isn’t a viable alternative.

Whereas Google works by matching queries with relevant websites found and indexed by its crawlers and ranking them on their popularity and other metrics, Perplexity leverages the capabilities of an advanced large language model to understand the user’s query and its context. This approach has a long ways to go!

Who is Backing Perplexity?

Oddly the backers are mostly so far from major BigTech firms including Yann LeCun, the chief AI scientist for Meta Platforms Inc. Its backers also include at least six current and former AI researchers at Google and sister company DeepMind, including Jeff Dean, Google’s senior vice president for research and AI, who was an angel investor in the company in its early stages, according to data from research firm PitchBook. 

It appears that individuals at Google itself are the main preseed and Series A backers! Among its investors are a striking number of Google’s own AI researchers. So this for me disqualifies this as being a real startup in a sense, the CEO does have a history at DeepMind as a Research intern and even OpenAI according to his LinkedIn profile.

Google is already fated to invest heavily in Anthropic, Character.AI and others. Generative A.I. is consolidating too fast already to have the commercial freedom to innovate a way that will be independent from the magnificent seven, which makes the Gen AI hype train a bit worrisome for the future of the internet. It’s actually a bit surprising and perplexing.

Angels in High Places - AI Bro Support

• Ashish Vaswani, one of the inventors of so-called Transformer models that inspired OpenAI’s GPT-4

• Jeff Dean, lead researcher at Google LLC’s AI unit

• OpenAI co-founder Andrej Karpathy

• Early AI angel investors Nat Friedman and Elad Gil.

So it’s backed by Silicon Valley AI and Angel royalty.

Open-Source Innovation

Still it’s pretty damn impressive that Perplexity is releasing its own AI large language models (LLMs) — pplx-7b-online and pplx-70b-online, named for their parameter sizes, 7 billion and 70 billion respectively.

How they are leveraging Llama-2, Mistral and Amazon is super interesting.

This startup will likely need hundreds of millions of dollars of funding to even build a viable prototype to how Search works today. Can it reach that point?

Join their growing Discord Community!

In addition, Perplexity employs various techniques to maximize factual accuracy and minimize the generation of false information.

Their LLMs are fine-tuned and augmented versions of the open source mistral-7b and llama2-70b models from Mistral and Meta.

Conclusion

If you consider their actual budget in 2023, they have done amazing even to still be around.

How Gemini impacts Google Search will be something to watch, even as Microsoft’s flurry of Copilots might not improve search substantially at all.

Srinivas explained that LLMs will eventually change the way people interact with their computers to find and consume information.

Perplexity’s new LLMs are notable because, in addition to being available for other organizations to use and build their own apps upon through Perplexity’s API (application programming interface), they also aim to offer “helpful, factual, and up-to-date information”. But how many organizations will want to use their API? I’m not so clear on this.

Perplexity is actively building their own web index to make search results even more accurate and relevant. This could shake up the search engine market by offering better ways to find information. With a bit of luck and more funding, they might find a business model that enables them to actually survive.

I cannot say the same for the UK’s Stability A.I. that has a leadership and funding crisis. Because Search Advertising is so lucrative, even a Moonshot like Perplexity is potentially a very valuable startup. Because it’s so rare for anyone to take on Google, the AI community is betting on this startup to at least translate RAG and LLMs into a new kind of search experience.

Subscribe now

Hey Everyone,

With all the drama surrounding Sam Altman right now in November, 2023, I’d venture to guess that the most important person in AI is actually former Google CEO Eric Schmidt. From policy, to venture capital to other big bets.

It’s turning out that Aleph Alpha is more like the Anthropic of Europe and Mistral seems to be more on the open-source side of things. What will Kyutai evolve into? While I’m not exactly proud of how Eric Schmidt has turned out, he is throwing his weight and fortune around in some interesting ways.

Sovereign AI is still just a Dream

The French are an idealistic bunch, but are Billionaires idealists too? Billionaires Xavier Niel, Rodolphe Saadé and Eric Schmidt announced a new nonprofit artificial intelligence research lab in Paris, marking France’s latest push to develop sovereign AI technology. 

Buried beneath the OpenAI hype and drama, in Europe something was occuring that was much more interesting.

If the U.S. is centralizing A.I. without real regulation under the control of BigTech, in Europe you have a counter culture, they are celebrating open-source more Hugging Face style. It has a french vibe.

Kyutai is a French AI research lab with a $330 million budget that will make everything open source

But look at the phrasing of their word choice:

Visit Kyutai Website

OpenAI started out with idealistic foundations as well, and just look at how it turned out.

Sounds familiar eh! But what with OpenAI with European characteristics? If the U.S. has taken LLMs to the corporate cloud, moat and locked, what can the world do to counter that A.I. weaponization of greed?

A private, non-profit initiative

Kyutai was founded jointly by the iliad Group, the CMA CGM Group and Schmidt Futures.

• To achieve this, Kyutai will utilize the computing power provided by Scaleway, a subsidiary of the iliad Group, which boasts the highest-performance computing power for AI applications in Europe, the release said.

• It will work with PhD students, postdocs and researchers on research papers and open source projects. When Iliad originally unveiled this research lab, the firm said that Niel was committing €100 million to this project ($109 million at today’s exchange rate).

My prediction is this becomes the ‘DeepMind of Meta’, Meta also likes to pretend it’s really into supporting open-source. Watch Facebook fund this in the near future in 2024.

Since the UK has been relatively disappointing in AI startups in Generative A.I., France is looking half decent with some big attempts like Mistral, Hugging Face adn now Kyutai.

According to Xavier Niel, chairman and founder of the iliad Group, Europe has the potential to excel in the AI race, and creating an AI open research lab in Paris is crucial to leverage the expertise of French researchers, available computing power and the dynamic AI ecosystem.

Billionaires for Science

Three billionaires, Eric Schmidt, Xavier Niel, and Rodolphe Saadé, announced the launch of the lab.

The main reason Eric Schmidt is important, is his emphasis on AI in the development of Quantum computing and Science. He was important in the Biden Administration banning A.I. chips to China. Way more important than Sam Altman in the grand scope of things, sorry for the AI bros reading this.

What is their Mission?

“OUR MISSION IS TO
BUILD AND DEMOCRATIZE
ARTIFICIAL GENERAL INTELLIGENCE
THROUGH OPEN SCIENCE”

I have no idea what that really means, but three things are important here, that they name AGI and Open and Science.

Europe will try to donate its way to open-competition. TechCrunch says that this is just a starting point, as Kyutai is open to more donations. “What’s interesting with so many journalists in the room is that the project will potentially interest other investors,” Saadé said at a press conference after the announcement.

For Billionaires, they want to buy a better legacy. They want their wealth to matter and translate into the future in a sustainable way.

Kyutai Got the Chips

As Kyutai will work on foundational models, they will also need some compute power. The good news is that Scaleway, the cloud division of Iliad, recently acquired a thousand Nvidia H100 GPUs. These top-of-the-line GPUs are essential for inference and model training and will be available at cost for Kyutai.

Implications for the Future

What if we had foundational models geared for science? Not Elon Musk’s version of what means, but Europe’s? The sort of AI that could make AI research laboratories around AGI.

The creation of Kyutai is not only a milestone in France’s tech journey but also a testament to the growing importance of AI in various industries. With its strong scientific team and focus on open-source research, the lab is well-positioned to contribute to foundational models and projects in AI. The founders hope that the lab’s establishment will not be impeded by regulatory constraints, thus, facilitating unhindered technological progress in the realm of artificial intelligence.

Eric Schmidt is backing Future House, among many similar projects. The biggest impact of LLMs is likely going to be in science and innovation itself, in how R&D occurs.

Schmidt has an estimated net worth of $24.5 billion, according to the Bloomberg Billionaires Index.

If Google were a good company, which it is not, its Search Advertising revenue would mean something benevolent to the future. People like Eric Schmidt have leverage of how the internet got started and are the types of people who profiteered from Silicon Valley’s centralization around early winners like Google, who up until recently, was perhaps the dominant “pure-play AI” company of the world. But a lot has changed since that 2017 Google paper. All the researchers are gone, a lot of the talent has left soon after.

We need to rebuild what AI could be, and might become with a bunch of smaller bets on things like AI and Science.

This might be the legacy of Schmidt after all. Nobody is more responsible for the U.S. vs. China momentum in AI. For good or ill.

Credentials, Plans and Building

• Kyutai has already started hiring for its core scientific team. Six men took the stage this morning to talk about their previous work and what they have in mind for the research lab — Patrick Perez, Edouard Grave, Hervé Jegou, Laurent Mazaré, Neil Zeghidour and Alexandre Defossez.

• They previously worked for Meta’s AI research team FAIR, Google’s DeepMind division, Inria, etc.

• The Facebook-Meta connection is actually thing:

1. Patrick Perez, who previously worked for Valeo, is going to be the director of the research lab.

2. Kyutai has also put together a team of scientific advisors who are well-known AI researchers — Yejin Choi, Yann LeCun and Bernhard Schölkopf. They will just check everyone’s work once or twice a year and give feedback.

Eric is nearly 70, his productive years are nearly behind him. He tried very hard in the last decade to make his mark.

He’s funneled some of that fortune into philanthropic efforts like Schmidt Futures, an initiative that funds science and technology entrepreneurs. In recent months, he’s emerged as an influential voice on AI policy in Washington. He certainly tried to be a mover and shaker in ways that even the media has missed.

A media that has tried to keep up with him.

Follow Kyutai on X here.

• One of the reasons Kyutai thinks it can convince some researchers to join its lab is that researchers will be able to publish research papers.

• Speaking at ai-Pulse, an event organised by Niel in Paris’s Station F startup campus on Friday, Schmidt said he backed Kyutai because he believes France can become an AI champion, as long as it can keep regulators in Brussels “under some level of control”.

So there’s also an element of lobbying here, to push America’s agenda of a lack of regulation on to Europe, to make France a bit less like Germany in this regard. The battle over A.I. regulation is deeper than it seems.

Support from the French Government

French President Emmanuel Macron and Digital Minister Jean-Noel Barrot have expressed their support for this initiative that seeks to prevent Europe from falling behind in AI.

The effort could lay the groundwork for an artificial intelligence industry in Europe. Although honestly Aleph Alpha and Mistral are doing that now.

The top UK Generative A.I. startup is likely Stability AI, that has been fraught with controversy too. It had so much trouble getting funding, it has to go with Intel in the end. Very bizarre.

Europe has been left behind more or less even as China has been hobbled by poor timing and geopolitical and economic downturns to keep up with the U.S. in the all important Generative A.I. front.

Rodriques, a biotechnology inventor who studied at the Massachusetts Institute of Technology, said Schmidt will fund Future House for its first five years. FutureHouse is a bit simliar. It’s a nonprofit that’s focused on building an artificial intelligence-powered assistant for the laboratory, with the lofty goal of overhauling the scientific research process. Read more.

Basically as OpenAI has gone commercial and even Anthropic has sold out to the Cloud in terms of Amazon and Google, more organizations are cloning the model of the old OpenAI, before it turned all corporate rogue under Sam Altman, under the watchful eyes of Satya Nadella with his chariots of Copilots. Which in a few years’ time might sound very lame, as lame as Cortana became in the last round of chatbot history.

Kyutai, FutureHouse and even X.AI seek AI at the intersection of science and human progress. There will be many others.

But Schmidt is a bit anti-European and pro-American. For example:

The EU AI Act

Schmidt — who is well-known for his criticism of EU regulators — pointed to the AI Act, which is currently being negotiated in Brussels, as an example of laws that might break innovation. But innovation does not occur at the service of a profiteering Cloud either with U.S. boardrooms and executives?

OpenAI, as the name still indicates, started as a nonprofit. But things changed drastically after Sam Altman started working full time on OpenAI in 2019. Perhaps OpenAI’s board relegalized the error years too late. Now it’s more like the DeepMind of Microsoft, or a wing of Microsoft Research focused on commercialization and poaching Google talent, basically serving Azure.

It’s near the end of 2023 that you realize how political AI has become. How fragmented and so much about power. The power of nations, the egos of billionaires, the dogmas of corporations.

Does Kyutai think it has enough H100s though? Iliad had acquired 1,000 Nvidia H100 GPUs that would be available to Kyutai to power the research of the new outfit. Compute, funding, competition - all in the name of ideals and science. What could possibly go wrong?

Hey Everyone,

For the developers and AI enthusiasts in the room, this 1st ever DevDay by OpenAI was quite a spectacle.

So OpenAI’s big day came and went. This is just a highly skimmable version in case you want a brief summary.

Subscribe now

Poll

🚨 Major new models and features release alert from OpenAI including new GPT4 Turbo model, lower prices, new Whisper and TTS model, inbuilt json support and more! 👇

Here are the new models and products:

🔹GPT-4 with 128k context window (300 pages in a single prompt).
🔹GPT-4 Turbo can accept images as inputs.
🔹Update on Speech to Text model.
🔹New Text-to-speech capability.
🔹GPT-4 finetuning.
🔹DALL·E 3 API.
🔹Lower price.

What’s New?

You can try the Assistants API beta without writing any code by heading to the Assistants playground.

New models and developer products announced at DevDay

Read the Blog

• OpenAI claims 2 million devs work on their product.

• OpenAI claims 100 million “weekly users” of ChatGPT

Video of Opening Keynote

Where are all the Devs?

Curiously the live-stream on YouTube was only watched live by less than 150k people. Not sure where all these “millions” of users are. Sure, some watched it on their website. Still, not the numbers you’d expect from a so-called viral product at the bleeding edge of Generative AI models and experience?

In a Nutshell

GPT-4 Turbo, currently in preview, will be three times cheaper for developers.

GPT-4 Turbo 128k

(1) GPT-4 Turbo 128k
OpenAI releases a new model that is GPT-4 level performance, but cheaper and faster. It has a 128k context window (300 pages book). Claims it is even accurate in longer context conversations.

(2) Substantial pricing reductions
The new GPT-4 Turbo will cost 2.75x cheaper on average, opening new opportunities and business models that didn’t make sense before.
GPT-3.5 Turbo 16k is also getting a price cut - 3x less than the previous price.

(3) New modalities
OpenAI releases its first text2Speech model, directly competing with the likes of ElevenLabs and PlayHT.
It also opens API access to DALL-E 3, GPT-4 (Turbo and Vision), and this new Text2Speech model.
Lastly, OpenAI will release a new Whisper V3.

(4) Better world knowledge
GPT’s knowledge cutoff was updated to April 2023 and will continue to improve over time.
Sam also hinted at allowing users to upload their own data, directly competing with RAG providers.

(5) More control
Improvements to function calling (better and support multiple functions in one call) and a special JSON mode for LLM builders using JSONs.
Also, users can now use a seed parameter to ensure the same prompt yields the same output (great to reduce hallucinations!).

(6) Customization
Users will be able to fine-tune GPT-4, just a few months after GPT-3.5 fine-tuning was released.
OpenAI will also work with customers in a Palantir-like model, landing its own researchers and engineers to work with companies to build well-crafted models, from preprocessing to training and deployment.

(7) Higher rate limits
2x tokens per minute with the ability to ask for rate increases via the account management interface.

(8) Copyright shield
Same as Google and Microsoft, OpenAI will defend any copyright infringement claims against its users.

OpenAI is launching a GPT store later this month.

The store will let users share and sell their custom GPT bots. Sam Altman says OpenAI is going to “pay people who build the most useful and the most used GPTs” a portion of the company’s revenue.

• Introducing GPTs.

OpenAI Tries to Build Momentum on ChatGPT Hyped Year

• Altman also shared today that over two million developers use the platform, including more than 92 percent of Fortune 500 companies.

• The company announced that DALL-E 3, OpenAI’s text-to-image model, is now available via an API after first coming to ChatGPT and Bing Chat.

Try ChatGPT

• OpenAI is launching a new API, called the Assistants API, that will help developers build “agent-like experiences” within their apps.

GPTs can be made with no coding experience, and can be as simple or complex as you like, as they explained.

What are GPTs by OpenAI?

The company said they are rolling out custom versions of ChatGPT that you can create for a specific purpose—called GPTs. GPTs are a new way for anyone to create a tailored version of ChatGPT to be more helpful in their daily life, at specific tasks, at work, or at home—and then share that creation with others.

• OpenAI claimed 460 companies on the Fortune 500 companies are using its products.

• GPTs will be available to paying ChatGPT Plus subscribers and OpenAI enterprise customers, who can make internal-only GPTs for their employees.

• GPTs are perhaps the most intriguing new feature they announced for me. We know that Character.AI has been very successful especially with GenZ and younger users.

Get more details of the below here.

Function calling updates

Function calling lets you describe functions of your app or external APIs to models, and have the model intelligently choose to output a JSON object containing arguments to call those functions.

Improved instruction following and JSON mode

GPT-4 Turbo performs better than our previous models on tasks that require the careful following of instructions, such as generating specific formats (e.g., “always respond in XML”). It also supports our new JSON mode, which ensures the model will respond with valid JSON.

Reproducible outputs and log probabilities

The new seed parameter enables reproducible outputs by making the model return consistent completions most of the time.

Learn more about X.AI’s Grok chatbot here.

Assistants API, Retrieval, and Code Interpreter

An assistant is a purpose-built AI that has specific instructions, leverages extra knowledge, and can call models and tools to perform tasks. The new Assistants API provides new capabilities such as Code Interpreter and Retrieval as well as function calling to handle a lot of the heavy lifting that you previously had to do yourself and enable you to build high-quality AI apps.

Assistants also have access to call new tools as needed, including:

• Code Interpreter: writes and runs Python code in a sandboxed execution environment, and can generate graphs and charts, and process files with diverse data and formatting. It allows your assistants to run code iteratively to solve challenging code and math problems, and more.

• Retrieval: augments the assistant with knowledge from outside our models, such as proprietary domain data, product information or documents provided by your users. This means you don’t need to compute and store embeddings for your documents, or implement chunking and search algorithms. The Assistants API optimizes what retrieval technique to use based on our experience building knowledge retrieval in ChatGPT.

• Function calling: enables assistants to invoke functions you define and incorporate the function response in their messages.

Longer Context Windows

• GPT-4 Turbo will offer a longer context window that can handle 300 pages of information.

Will OpenAI be the Apple of GPTs and LLMs?

The App Store model has proven unbelievably lucrative for Apple where they also plan to take a 30% cut, so it should come as no surprise that OpenAI is attempting to replicate it here, literally cloning it, but does it make sense? Not only will GPTs be hosted and developed on OpenAI platforms, but they will also be promoted and evaluated.

Certainly this will lead to a lot of products built on top of OpenAI that may or may not have much longevity.

It definitely will have a high grifting-factor a la Sammy:

“We’re going to pay people who make the most used and most useful GPTs with a portion of our revenue,” and they’re “excited to share more information soon,” Altman said.”

It’s difficult to imagine this taking off in the way OpenAI hopes.

OpenAI is frantically trying to build a business model that has an in-built moat to its first mover advantage, even as other similar companies gain more funding.

OpenAI has done heavy marketing that it’s an AGI company, though Sam Altman claims ChatGPT went viral by word of mouth alone. If Twitter bots and paid influencers are word of mouth, I guess that means that Generative A.I. will save the world as well!

Still the concept of GPTs, a Store and Turbo are an impressive trinity of announcements. Which we mostly already knew about.

GPT-4 Turbo will also “see” more data, with a 128K context window, which OpenAI says is “equivalent to more than 300 pages of text in a single prompt.”

The event was extremely commercial and sort of making OpenAI seem to be more mainstream than they are. As far as the public knows, they are one year old company, but have been around since December, 2015.

OpenAI says it will be monitoring activity of GPTs to block things like fraud, hate speech, and “adult themes.” As TechCrunch notes, it’s not clear at this point whether there will be the ability to simply charge for your GPT, or whether it will be strictly revenue sharing.

• Since Microsoft makes 75% of OpenAI’s revenue until their deal is paid off, Microsoft appears to be giving OpenAI the front seat, while Microsoft are the puppet masters from the back room.

• OpenAI has a facade of being very privacy and ethics conscious. When the GPT Store launches down the road, OpenAI will only accept agents from people who have verified their identity. Initially, GPTs will be accessible through shareable web links. Still for me personally I don’t find Sam Altman a very trustworthy source. The “grift” factor is really high here.

The company still has that AGI method to its corporate madness of close-sourced commercialization of LLMs. Calmly reminding us that ultimately, OpenAI sees its GPT platform as bringing it one step closer to its main goal: the creation of an AI superintelligence, or AGI.

View OpenAI’s Tweet about GPTs. - Go to X post.

https://twitter.com/OpenAI/status/1721594380669342171

3 Main Takeaways

1. Their new model GPT-4 Turbo supports 128K context and has fresher knowledge than GPT-4. Its input and output tokens are respectively 3× and 2× less expensive than GPT-4. It’s available now to all developers in preview.

2. Assistants API and new tools (Retrieval, Code Interpreter) will help developers build world-class AI assistants within their own apps.

3. The platform is becoming multimodal. GPT-4 Turbo with Vision, DALL·E 3, and text-to-speech are all now available to developers.

Kudos to OpenAI for trying so hard.

That’s all I got.

Share

Hey Everyone,

I’ve been really impressed by the startup Anyscale, who have recently announced Anyscale Endpoints - fine-tuning.

Learn More

Anyscale Endpoints supports fine-tuning

If you think about how this is getting cheaper in the Open-source space, it’s going to be pretty astounding how accessible this becomes in the 2020s.

Their LLM fine-tuning API. Fine-tuning is essential for cost reduction. Fine-tuning can enable task-specific performance superior to GPT-4 at 1/300th the cost.

Anyscale *Private* Endpoints

Private Endpoints let's you deploy the entire backend of the LLM API (including the models) *in your cloud*, *private* for your business, and 100% *customizable*. This is a unique product on the market.

$1 per million tokens

Anyscale Endpoints charges $1 / million tokens for Llama-2-70B (and less for other models). We optimize every layer of the stack to make this possible. This is the *lowest price point* for Llama-2-70B.

These offerings build on everything they have been developing with Ray over the past few years as well as all of the advances in the open source community.

Anyscale’s goal is to provide the most performant and cost-efficient infrastructure for open LLMs.

Anyscale built a n pen-source distributed computing platform Ray, and is now launching a new service called Anyscale Endpoints and their new API, that it says will help developers integrate fast, cost-effective and scalable large language models into their applications. I got to admit, I find this really quite impressive.

It has to be more affordable and while companies like OpenAI are charging too much without adequate control, customization, privacy and cost efficiency, others are building stuff that can offer those things.

With the launch of Anyscale Endpoints, developers now have a simple way to build distributed applications that leverage the most advanced generative AI capabilities using the application programming interfaces of popular LLMs such as OpenAI LP’s GPT-4.

I’m a huge fan of what teams like Anyscale, Cohere, AI21 Labs and others are doing in the space.

Here's a quick glimpse at what Anyscale Endpoints offers:

• State-of-the-art open-source and proprietary performance and cost optimizations.

• A serverless approach to running open-source LLM models, mitigating infrastructure complexities.

• A seamless transition for running base or fine-tuned models on your cloud.

• Streaming response.

• Utilization of the power of Ray Serve and Ray Train libraries.

It’s really something I hope startups and smaller enterprise and SMBs can experiment with.

Additional services from the Anyscale Platform:

• If you need to fine-tune and deploy models on Anyscale's cloud infrastructure, our platform offers a comprehensive suite of services.

• For a more secure, customized environment, the Anyscale platform provides the necessary resources and support for fine-tuning and deploying LLMs in your own cloud managed by Anyscale's cloud infrastructure.

Share

Hey Everyone,

So today July 18th, 2023 something impactful was announced.

Subscribe now

Yann LeCun on LinkedIn posted:

This is going to change the landscape of the LLM market.
Llama-v2 is available on Microsoft Azure and will be available on AWS, Hugging Face and other providers

Pretrained and fine-tuned models are available with 7B, 13B and 70B parameters.

Llama-2 website: https://ai.meta.com/llama/
Llama-2 paper: https://lnkd.in/djeTCvfn

A number of personalities from industry and academia have endorsed our open source approach: https://lnkd.in/dSge3PGV

Hugging Face aficionados reported on it:

Interconnects

LLAMA 2: an incredible open-source LLM

Programing note: we're coming early because this is one of the biggest AI stories of the year. If this email is cut off, please read on the web…

Read more

3 years ago · Nathan Lambert

What does it all mean?

• Llama 2 is free for research and commercial use. 

• Microsoft and Meta are expanding their long-standing partnership, with Microsoft as the preferred partner for Llama 2.

• They’re opening access to Llama 2 with the support of a broad set of companies and people across tech, academia, and policy who also believe in an open innovation approach to today’s AI technologies.

• Meta AI is also committed to building responsibly and are providing resources to help those who use Llama 2 do so too.

Recently the AI arm of Mark Zuckerberg owned Meta, has introduced a multimodal model  “CM3leon” (pronounced like chameleon), that does both text-to-image and image-to-text generation.

Meta AI has been more active in the A.I. space in recent weeks than many presumed they would be.

Microsoft was boasting with PR about the partnership.

Today, at Microsoft Inspire, Meta and Microsoft announced support for the Llama 2 family of large language models (LLMs) on Azure and Windows. Llama 2 is designed to enable developers and organizations to build generative AI-powered tools and experiences. Meta and Microsoft share a commitment to democratizing AI and its benefits and we are excited that Meta is taking an open approach with Llama 2. We offer developers choice in the types of models they build on, supporting open and frontier models and are thrilled to be Meta’s preferred partner as they release their new version of Llama 2 to commercial customers for the first time.

Microsoft’s $13 Billion in OpenAI, did anything but help democratize A.I. to be honest. That BigTech is pretending they are behind open-source innovation is getting a bit weird.

It’s as if Azure is trying to take credit? Very strange. The PR adds: “Expanding Azure AI model catalog and Windows availability. Llama 2 is the latest addition to our growing Azure AI model catalog. The model catalog, currently in public preview, serves as a hub of foundation models and empowers developers and machine learning (ML) professionals to easily discover, evaluate, customize and deploy pre-built large AI models at scale.”

Meta is boasting about CM3leon and now this. BigTech a tyranny of PR, now somehow admitting open-source is important.

Read Meta's Blog About Llama 2

Listen that the announcement, is coupled with details of Microsoft’s pricing plans for artificial intelligence in some software, underscores rapid AI development at big tech companies and the sad monopolization of the state of Generative A.I.

Meta AI the leader in open-source LLMs? Microsoft being agonistic? Hardly.

Read the Paper

• Llama 2 is available for free for research and commercial use.

• Over 100,000 requested access to Llama 1, Meta Says.

The paper is 76 slides long.

• Llama 2 was trained on 40% more data.

The benefits of open-source have BigTech worried, so they are pretending they are a big part of the wave.

Open-source models certainly have a mix of benefits and negatives. Providing a wide group of people with an AI's differing capabilities can help the model learn faster, given the increased data it's receiving. Additional eyes can also spot bugs and security issues, giving developers a chance to fix these problems sooner.

Each model trying to out-do the others. So this is what it comes to.

According to Meta, LLaMa 2 was trained on 40 percent more data when compared to LLaMa 1, which includes information from “publicly available online data sources.” It also says it “outperforms” other LLMs like Falcon and MPT when it comes to reasoning, coding, proficiency, and knowledge tests.

Benchmarks

Llama-2 looks good on paper.

BigTech keeps trying to hype up what LLMs can do. Large Language Models (LLMs) have shown great promise as highly capable AI assistants that excel in complex reasoning tasks requiring expert knowledge across a wide range of fields, including in specialized domains such as programming and creative writing.

GitHub

https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md

Read Responsible Use Guide

These guys seem to take A.I. models very seriously.

In opening up LLaMA, Meta said it wants to improve safety and transparency. Meta AI ain’t no Anthropic, I can tell you that. Google has a 10% stake in Anthropic A.I.

The company said the LLaMA 2 model has been “red-teamed,” or tested for safety by “generating adversarial prompts to facilitate model fine-tuning,” both internally and externally.

Safter is Better

Meta AI branding themselves as the good guys in A.I. innovation.

Subscribe now

Download the Model

Hey Everyone,

I won’t be posting as much during the summer months, because a lot of folk are simply with family, on holiday, traveling and in digital detox mode. I’ll still be posting, just not my usual frantic pace.

Eleven Labs is one of my favorite voice cloning tools, here’s how simple it is to use:

Referrals on this Newsletter unlock free paid complementary subscriptions. Thanks for sharing!

Refer a friend

I’ve been tracking A.I. startups in their Venture Capital funding very closely since the Generative A.I. I find this one just has an “it” factor that’s hard to explain vs. its competitors.

The participants of the funding round also really interested me. Voice cloning will quickly become mainstream and change how we interact with A.I. assistants in the future.

On June 20th, 2023 the startup announced the closure of a $19 million Series A round co-led by entrepreneurs Nat Friedman and Daniel Gross alongside Andreessen Horowitz (better known as a16z). Nearly exactly four years ago a16z did a Series A with this platform, Substack.

Other participants included heavyweights Creator Ventures, SV Angel, Instagram co-founder Mike Krieger, Oculus co-founder Brendan Iribe, Deepmind and Inflection AI co-founder Mustafa Suleyman and O’Reilly Media founder Tim O’Reilly.

A lot of Generative A.I. startups just won’t make it, since there won’t be mainstream demand for their services. ElevenLabs is not one of those.

Subscribe now

Read more

Hello Everyone,

In a nutshell, I think MusicGen is pretty cool.

Will A.I. augment creativity?

Felix Kreuk, an AI research engineer at Meta, showed off the “MusicGen’s” capabilities in a Twitter thread late last week.

Meta A.I. is doing some really neat work on open-source A.I.  All credit to Jade Copet, Felix Kreuk, Tal Remez, Itai Gat, David Kant, Gabriel Synnaeve and Yossi Adi.

The sample posted here on LinkedIn recently, blew me away.

It made me laugh hearing the Palantir CEO talk about LLMs as poetry-generators.

The LLM-poetry quote happens at: 2:45 of this video. I don’t know why, but it makes me laugh.

Palantir AIP looks pretty dangerous frankly. But seriously, AI has managed to intrude on most artistic endeavors, and now it’s fully come for the music industry perhaps.

Speaking about poetry, how about music eh?

Text to Music is here in 2023

• Listen to our samples: https://ai.honu.io/papers/musicgen/

• Read the paper: https://arxiv.org/abs/2306.05284

• Play with our demo on HuggingFace: https://huggingface.co/spaces/facebook/MusicGen

Read the Paper

MusicGen

Meta's Audiocraft research team has just released MusicGen, an open source deep learning language model that can generate new music based on text prompts and even be aligned to an existing song.

See the Tweet of Felix here.

MusicGen is basically a simple and controllable music generation model. MusicGen can be prompted by both text and melody. They release code (MIT) and models (CC-BY NC) for open research, reproducibility, and for the music community:

AudioCraft on GitHub

Subscribe now

Read more

Welcome Back!,

There are ways LLMs will impact search that we aren’t thinking very well about. Shay Banon created the precursor to Elasticsearch, called Compass, in 2004. Nearly 20 years later ChatGPT and ElasticSearch are compatible. That is, you can connect ChatGPT to proprietary data stores using Elasticsearch and build question/answer capabilities for your data.

How LLMs leverage proprietary data across industries is a holy grail for some products. We are seeing this already with a proliferation of LLMs specific to industries and the specialized high quality data and Big Data of some companies.

Share

It’s actually pretty easy:

• Elasticsearch is a distributed, free and open search and analytics engine for all types of data, including textual, numerical, geospatial, structured, and unstructured. Elasticsearch is built on Apache Lucene and was first released in 2010 by Elasticsearch N.V. (now known as Elastic).

• Known for its simple REST APIs, distributed nature, speed, and scalability, Elasticsearch is the central component of the Elastic Stack, a set of free and open tools for data ingestion, enrichment, storage, analysis, and visualization. Commonly referred to as the ELK Stack (after Elasticsearch, Logstash, and Kibana), the Elastic Stack now includes a rich collection of lightweight shipping agents known as Beats for sending data to Elasticsearch.

Even from a non-technical point of view, I’ve been interested in Elasticsearch for some time and new kinds of search potential regarding enterprise data.

Cleary ChatGPT and Elasticsearch has incredible synergy.

1. Elasticsearch lets you perform and combine many types of searches — structured, unstructured, geo, metric — any way you want. Start simple with one question and see where it takes you.

2. It's one thing to find the 10 best documents to match your query. But how do you make sense of, say, a billion log lines? Elasticsearch aggregations let you zoom out to explore trends and patterns in your data.

3. Now imagine all of this made sense in a ChatGPT like interface with those GPT-4 capabilities?

The Human Machine Interface

Whatever you think of Generative A.I., the human-AI hybrid workforce is coming into being now.

Subscribe now

Let’s walk though some of the finer details:

Read more

Welcome Back!

I actually think Amazon’s Bedrock announcement is pretty major.

Amazon Bedrock provides you the flexibility to choose from a wide range of FMs built by leading AI startups and Amazon so you can find the model that is best suited for what you are trying to get done.

Share

Text generation

Create new pieces of original content, such as short stories, essays, social media posts, and web page copy.

Chatbots

Build conversational interfaces such as chatbots and virtual assistants to enhance the user experience for your customers.

Search

Search, find, and synthesize information to answer questions from a large corpus of data.

Text summarization

Get a summary of textual content such as articles, blog posts, books, and documents, to get the gist without having to read the full content.

Image generation

Create realistic and artistic images of various subjects, environments, and scenes from language prompts.

Personalization

Help customers find what they’re looking for with more relevant and contextual product recommendations than word matching.

So what’s going on here?

Generative A.I. is a Battle in the Cloud

If Microsoft got OpenAI and Google has Anthropic, who does Amazon have?

While Stability AI is a likely candidate in the future, for now it would seem that instead of partnering with some up-and-coming AI startup, Amazon is opening the floodgates for more AI integration on its cloud web service platform.

Amazon is also offering free access to its CodeWhisperer in an attempt to take on GitHub Copilot. Amazon is making its AI-powered coding assistant free for individual developers, undercutting the $10 per month pricing of its Microsoft-made rival. The rivalry between AWS and Azure is getting hot under the umbrella of Generative A.I. it would seem, even as Google has struggled to keep up with OpenAI in a credible way in recent months in early 2023.

Generative A.I. in this climate feels like a relentless, untamed wilderness, AI’s enigmatic potential stretches to infinity, but in reality is likely a bubble before a storm of a recession, but don’t tell the people competing in the Cloud.

Bedrock is Amazon’s most forceful play yet for the generative AI market, which could be worth close to $110 billion by 2030, according to estimates from Grand View Research.

Amazon CEO Andy Jassy says ‘really good’ A.I. models take ‘billions of dollars’ to train

Cloud leaders are using the hype around Generative A.I. startups to try to turn the tides in their favor, just like we are seeing with search Advertising and will see with other kinds of advertising.

Through Bedrock, clients can use language models from Amazon and startups AI21 and Anthropic, as well as Stability AI’s model for turning text into images. Amazon already appears to have big name clients and customers. Amazon said Pegasystems, Deloitte and Accenture and are among the companies looking forward to using Bedrock.

Cloud growth is slowing down in reality just like the Advertising slowdown, and A.I. hype is the shield of choice these days in the world of business.

Subscribe now

Let’s go into some more detail:

Read more

Hey Everyone,

The Cloud race of A.I. startup affiliating is heating up in 2023. AWS already got Stability.AI late last year, now it’s the turn of Hugging Face. This is actually a bigger deal than people realize.

Just as BingAI moved fast and is breaking things for Microsoft’s brand reputation with their $13 Billion partnership with OpenAI, Google Cloud and AWS are keeping up with Azure in their own ways.

Hugging Face is a huge hub for all things machine learning as you know. Amazon Web Services (AWS), the cloud computing arm of Amazon.com Inc, on Tuesday said it is collaborating with startup Hugging Face, a software development hub, to make it easier to carry out artificial intelligence work (AI) in Amazon's cloud.

Hugging Face is one of the most AI-native startups around. Think about it, Hugging Face, Inc. is an American company that develops tools for building applications using machine learning. It is most notable for its Transformers library built for natural language processing applications and its platform that allows users to share machine learning models and datasets.

Read the Announcement

Of course Hugging Face is using the “democratizing AI” card in how they pitch this. They said:

There have been significant advances in new Transformer and Diffuser machine learning models that process and generate text, audio, and images. However, most of these popular generative AI models are not publicly available, widening the gap of machine learning capabilities between the largest tech companies and everyone else.

Share Datascience Learning Center

If you go with a Cloud partner, it’s not all accessibility and democratization. But it sounds great on paper.

AWS and Hugging Face are partnering to contribute next-generation models to the global AI community and democratize machine learning. Through the strategic partnership, Hugging Face will leverage AWS as a preferred cloud provider so developers in Hugging Face’s community can access AWS’s state-of-the-art tools (e.g., Amazon SageMaker, AWS Trainium, AWS Inferentia) to train, fine-tune, and deploy models on AWS.

Subscribe now

Read more

Hey Everyone,

So ChatGPT was impressive for some individuals who wanted hacks to be more productive, I get that. But when many organizations and online communities are banning its use, it’s a full blown crisis of A.I. alignment.

That’s not of course a story the mainstream media wants or is good at covering.

As someone who believes in A.I. ethics, I don’t think unleashing something that isn’t ready too early into the wild is a great idea. Generative A.I. basically means a lot more spam and unreliable content. We’ve already been struggling with content moderation at scale, with A.I. systems that aren’t able to catch all the bad actors.

Subscribe now

Famously a lot of School boards have to re-write their rules around students using ChatGPT, as if a supposedly forthcoming ChatGPT watermark in the works can save them?

New York City public schools ban access to AI tool that could help students cheat

What you may not realize, it’s banning ChatGPT is becoming a bit of a trend in early 2023.

Top AI conference bans use of ChatGPT and AI language tools to write academic papers

One of the world’s most prestigious machine learning conferences has banned authors from using AI tools like ChatGPT to write scientific papers, triggering a debate about the role of AI-generated text in academia.

The International Conference on Machine Learning (ICML) announced the policy earlier this week, stating, “Papers that include text generated from a large-scale language model (LLM) such as ChatGPT are prohibited unless the produced text is presented as a part of the paper’s experimental analysis.”

Meanwhile real artists are also being banned for having a style that resembles the text-to-image art. The Art subreddit on Reddit has permanently banned on artist solely because their style looks similar to AI art, or at least that’s what one mod thinks.

Remember even back in December Stack Overflow had to ban ChatGPT, for the pollution and inaccuracies it was giving. While that story was pushed under the rug in 2022, now in 2023 we are seeing a lot more organizations, communities, online forums and institutions having to ban ChatGPT. Not because it’s amazing, but because the tool actually pollutes, infringes or devalues those communities and their goals.

So let’s not all be bullish at the same time for the emergence of GPT-4 and better RLHF Conversational A.I, we are a long ways from massive scale. There’s good reasons Google didn’t release LaMDA to the public, even when its employees claiming for fame or whatever that it was practically sentient. We are nowhere near AGI, and it’s important to project fact-verified realities and not just the hype spin.

Why do you suppose ChatGPT is needing to be banned in 2023? It’s just a limited demo stuck in 2021 data, just learning to optimize better with human feedback. It’s not the product that we’ll experience later this year on Bing. Most of the A.I. writers I see have been programmed to walk the hype via clickbait but I’m uncertain how much I want to jump on that bandwagon. Even as an indie journalist, I’m more interested in fair and balanced coverage and accountability journalism.

ChatGPT and Conversational A.I. is hard, it’s not simply a transformative product that is good for the internet and everyone who uses it. Sam Altman has the heart of a venture capitalist, he’s not interested in his product being perfect before its released, he’s interested in how much his company is valuated at.

It's just January 7th, but ChatGPT is getting banned a lot in 2023 already. What does it mean for Microsoft Bing and GPT-4? Deliverance or mistrust in the hype? I'm reading some interesting points of view on Reddit. With A.I. we have to be long-termism and not give in to the temptation to join the choir. You can be an optimistic and a realistic at the same time, they don’t have to be mutually exclusive.

On LinkedIn people are trying to rationalize how OpenAI might indeed be worth that much? It’s fun to throw your hat in with the Big boys club, I get it. But the truth also matters. One writer literally claimed that the size of GPT-4 was much higher than more rational estimates indicated, which created an entire wave of misinformation around GPT-4, so I had a front seat to all of this and it was all a bit shocking to me.

People assume innovation has some kind of defacto right to disrupt and pollute our communities or that we should accept the “spirit of innovation” (mainly just pushed by BigTech) as some kind of progress of the scientific paradigm, but that’s not all that’s actually going on.

A.I. researchers are actually interested in democratizing A.I. with projects like Bloom and others. Open’s DALL-E 2 weren’t as open as Midjourney or Stability.AI. ChatGPT it appears will be a Microsoft Bing product, supposedly enabling them to compete with Google. Google for itself now may have to launch their own demo of LaMDA (likely before it’s even ready for public consumption). Google executives were seriously worried about the reputational risk, this is a red flag for A.I. Safety.

Share

Read more

Hey Guys,

So I’ve been saying that RLHF was a very important key to how GPT-3 became GPT-3.5 and how ChatGPT and davinci-003 perform so well. They are just version of InstructGPT, or sisters we might say.

Reinforcement Learning with Human Feedback is the Key to Aligning Language Models to Follow Instructions

See my brief dive into RHLF here. For more info watch this video.

So InstructGPT was what resulted when OpenAI trained language models that are much better at following user intentions than GPT-3 while also making them more truthful and less toxic, using techniques developed through our alignment research. RHLF is the key to creating more aligned A.I.

These InstructGPT models, which are trained with humans in the loop, are now deployed as the default language models of their API. This was of January, 2022.

Read the Blog

Before ChatGPT and Davini-003 there was InstructGPT. The OpenAI API is powered by GPT-3 language models which can be coaxed to perform natural language tasks using carefully engineered text prompts.

This week, OpenAI open sourced Point-E, a machine learning system that creates a 3D object given a text prompt. According to a paper published alongside the code base, Point-E can produce 3D models in one to two minutes on a single Nvidia V100 GPU. So OpenAI is not standing still, and neither are Google on this.

What is InstructGPT?

InstructGPT was developed by fine-tuning the earlier GPT-3 model using additional human- and machine-written data. The new model had an improved ability to understand and follow instructions, and that’s what essentially made ChatGPT possible, which went viral about 7 months later.

Paper link

1. Making language models bigger does not inherently make them better at following a user's intent.

2. For example, large language models can generate outputs that are untruthful, toxic, or simply not helpful to the user. In other words, these models are not aligned with their users. In this paper, we show an avenue for aligning language models with user intent on a wide range of tasks by fine-tuning with human feedback.

3. Starting with a set of labeler-written prompts and prompts submitted through the OpenAI-API, we collect a dataset of labeler demonstrations of the desired model behavior, which we use to fine-tune GPT-3 using supervised learning. We then collect a dataset of rankings of model outputs, which we use to further fine-tune this supervised model using reinforcement learning from human feedback (RLHF).

4. We call the resulting models InstructGPT. In human evaluations on our prompt distribution, outputs from the 1.3B parameter InstructGPT model are preferred to outputs from the 175B GPT-3, despite having 100x fewer parameters.

5. Moreover, InstructGPT models show improvements in truthfulness and reductions in toxic output generation while having minimal performance regressions on public NLP datasets. Even though InstructGPT still makes simple mistakes, our results show that fine-tuning with human feedback is a promising direction for aligning language models with human intent.

Read the Paper

ChatGPT trained with GPT-4 may be able to for instance, allow Microsoft Bing to better compete with Google, or develop into language tutors.

Large language models (LMs) can be “prompted” to perform a range of natural language processing (NLP) tasks, given some examples of the task as input.

It’s actually InstructGPT that was a breakthrough in quality as rated by human feedback.

InstructGPT models (PPO-ptx) as well as its variant trained without pretraining mix (PPO) significantly outperform the GPT-3 baselines. This was already with InstructGPT, nearly a year ago already as of December, 2022.

As impressive as davini-003 is (Scale.AI), I’d argue that InstructGPT was perhaps the breakthrough that was most important.

Like Matt Bastian of the Decoder writes:

The new GPT-3 model “text-davinci-003” is based on the InstructGPT models introduced by OpenAI earlier this year, which are optimized with human feedback. These models have already shown that AI models trained with RLHF (Reinforcement Learning from Human Feedback) can achieve better results with the same or even lower parameters.

Specifically, OpenAI used reinforcement learning from human feedback (RLHF; Christiano et al., 2017; Stiennon et al., 2020) to fine-tune GPT-3 to follow a broad class of written instructions. Davinci-003 made it far better with longer texts, thus in a sense making the ChatGPT we know today possible.

Share

Some of the innovations/or tools related to coding have been interesting:

The key of InstructGPT is how OpenAI collected a dataset of human-written demonstrations of the desired output behavior on (mostly English) prompts submitted to the OpenAI API3 and some labeler-written prompts, and use this to train their supervised learning baselines.

Now in 2023, we’ll be able to train GPT-4 and whatever else with better RLHF, companies like Google, ByteDance and Microsoft will be in a race to embed this technology into their products to make them even smarter.

Please understand how this is going mainstream even in 2023 or 2024.

Read more

Hey Guys,

Just a quick note, Meta is releasing PyTorch 2.0.

PyTorch Important for Machine Learning

As you know, among the most widely used machine learning (ML) technologies today is the open-source PyTorch framework.

It’s becoming pretty dominant. PyTorch got its start at Facebook in 2016 with the 1.0 release debuting in 2018. It’s come pretty far in just five years.

In September 2022, Meta moved the PyTorch project to the new PyTorch Foundation, which is operated by the Linux Foundation. Today, December 2nd, 2022, PyTorch developers took the next major step forward for PyTorch, announcing the first experimental release of PyTorch 2.0.

So it looks like the new release promises to help accelerate ML training and development, while still maintaining backward-compatibility with existing PyTorch application code.

Read the Blog

To complement the PyTorch 2.0 announcement and conference, they have also posted a comprehensive introduction and technical overview within the Get Started menu at https://pytorch.org/get-started/pytorch-2.0.

This release is intended to improve performance speed as well as add support for Dynamic Shapes and Distributed while still maintaining the same eager-mode development and user experience.

They claim PyTorch’s biggest strength beyond their “amazing community” is that they continue as a first-class Python integration, imperative style, simplicity of the API and options.

PyTorch 2.0 Updates

PyTorch 2.0 introduces torch.compile, a compiled mode that accelerates your model without needing to change your model code. On 163 open-source models ranging across vision, NLP, and others, we found that using 2.0 speeds up training by 38-76%.

PyTorch 2.0 is *100%* backward-compatible.
The codebase is the same, the APIs are the same, and the way you write models is the same.
We are calling it 2.0 because it marks adding a significant new set of features.

New components in PyTorch 2.0:

- TorchDynamo generates FX Graphs from bytecode analysis
- AOTAutograd generates backward graphs ahead-of-time
- PrimTorch introduces a small operator set to make backends easier 
- TorchInductor: a DL Compiler powered by OpenAI Triton

PyTorch 2.0 also introduces `torch.compile`, a new capability that improves PyTorch performance and starts the move for parts of PyTorch from C++ back into Python.

Additionally, several new technologies are included in the update, including: 

• TorchDynamo, which works to safely capture PyTorch programs using Python Frame Evaluation Hooks. 

• AOTAutograd to overload PyTorch’s autograd engine as a tracing autodiff for generating ahead-of-time backward traces.

• PrimTorch to canonicalize ~2000+ PyTorch operators down to ~250 primitive operators that developers can use to build a complete PyTorch backend. 

• TorchInductor is a deep learning compiler that generates code for several accelerators and backends. 

Hugging Face also got in on it.

PyTorch Conference Live-Stream Watch

• The most recent generally available release of PyTorch is version 1.13, which came out at the end of October, 2022.

• Chintala emphasized that now is the right time for PyTorch 2.0 because the project is introducing an additional new paradigm in the PyTorch user experience, called torch.compile, that brings solid speedups to users that weren’t possible in the default eager mode of PyTorch 1.0.

• He explained that on about 160 open-source models on which the PyTorch project validated early builds of 2.0, there has been a 43% speedup and they worked reliably with the one-line addition to the codebase. 

It’s vaguely interesting in that:

What do you guys think?

Leave a comment

As you know, PyTorch gets the first part of its name (Py) from the open-source Python programming language that is widely used in data science. It’s becoming one of the dominant tools in Machine Learning.

PyTorch is a machine learning framework based on the Torch library, used for applications such as computer vision and natural language processing, originally developed by Meta AI and now part of the Linux Foundation umbrella. It is free and open-source software released under the modified BSD license.

“PyTorch 2.0 embodies the future of deep learning frameworks”

Work on Python 2.0 will be ongoing for the next several months with general availability not expected until March 2023.

Thanks for reading!

Get get access to all articles don’t forget to subscribe.

Subscribe now

Hey Everyone,

I wanted to talk a bit about LLMs and how they are integrating in a new vision for the future of artificial intelligence.

With GPT-4 being announced soon, let’s take a look at large language models, sometimes called LLMs are and why they will be transformative.

In fact, the celebrated writer and future analyst in an intro to his GPT-4 article wrote:

It’s been 29 months since OpenAI launched GPT-2, its large-language model which demonstrated the power of transformers-based neural networks. GPT-2 impressed with the quality of its natural text generation. Its successor GPT-3, a bigger, more complex model, delivered even more powerful results. 

Since GPT-3 was released in 2020, we have witnessed a wave of new innovations and products built on it and similar models. It isn’t just about feeding a model a text prompt and having it spew out lots of plausible-sounding (but not necessarily true) copy. We’re witnessing the same prompt-based approach to images, movies and more. 

The best introduction online I could find to LLMs is actually by Co:here.

Given how language is key to how humanity progressed, it’s not surprising to find large language models proving important in A.I.’s development.

1. Discussion of GPT-1 paper (Improving Language Understanding by Generative Pre-training).

2. Discussion of GPT-2 paper (Language Models are unsupervised multitask learners) and its subsequent improvements over GPT-1.

3. Discussion of GPT-3 paper (Language models are few shot learners) and the improvements which have made it one of the most powerful models NLP has seen till date.

Many of expect GPT-4 to launch between February and June, 2023.

Sam Altman’s rants on Twitter about Startups are almost as painful to read as Reid Hoffman’s “Newsletter” on LinkedIn, and details of GPT-4 are murky at best.

The year was 2017, and the Transformer came on the scene. The transformer neural network is a novel architecture that aims to solve sequence-to-sequence tasks while handling long-range dependencies with ease. It was first proposed in the paper “Attention Is All You Need.” and is now a state-of-the-art technique in the field of NLP.

A transformer model is a neural network that learns context and thus meaning by tracking relationships in sequential data like the words in this sentence.

Transformer models apply an evolving set of mathematical techniques, called attention or self-attention, to detect subtle ways even distant data elements in a series influence and depend on each other.

First described in a 2017 paper from Google, transformers are among the newest and one of the most powerful classes of models invented to date. They’re driving a wave of advances in machine learning some have dubbed transformer AI.

Read more

Hey Guys,

After a bunch of creators are monetizing their content with GPT-4 rumors, I thought it might be interesting to survey said rumors.

One step towards the Machine Economy, let’s go!

Since it’s highly speculative, I’m only sending this out to the few supporters of this Newsletter.

Share

The day for the release of GPT-4 is getting closer.

A.I. anticipation is giving us a chance to ignore fact-checking and say something outladish on Twitter, Reddit, or Substack (for that matter). According to a new A.I. writer on Substack: In 2020, GPT-3 surprised everyone with a huge performance leap from GPT-2 and set unprecedented expectations for its successor.

Yes that’s usually the way things work.

However people are saying things on Twitter that are a bit amusing to follow.

Here’s what I found: (let me just say I have no idea who these Twitter folk are)

This is mostly for entertainment and I’m going to let the Tweets do the bragging:

Like human level AI next level?

METATREND #3: AI-HUMAN COLLABORATION WILL SKYROCKET ACROSS ALL PROFESSIONS

Reminds me of Peter Diamandis and the transhumanists at Singularity University.

Go Big or Go Home

Meta’s missteps over its Galactica LLM public demo, does make you wonder about how LLMs are used. (Which incidentally reminds me of the show Battlestar Galactica).

In July 2022, OpenAI launched DALLE2, a state-of-the-art text-to-image model. In early 2023, we’ll be due for GPT-4, my guess is in April, 2023.

Though will it really be that big?

The hype generation on Twitter and elsewhere is really getting out of hand. LinkedIn is just as bad for things like Quantum computing, which I cover. This sort of “rumour” amounts to pumping up OpenAI which amounts to pumping up Microsoft, their main sponsor.

Human Level AI Probably Won’t Save Humanity

Just don’t tell the people on Twitter that.

Yes powerful, but for whom?

A.I is summoning you to help, immediately!

So long as the robots tie their own shoe laces I’m okay with this:

A.I. Progress is Making Grown Men Emotional

I did not find significant rumors that help me understand how significant GPT-4 will be above GPT-3, in earth years, already 2.5 years later after GPT-3.

This article might help though.

Thanks for reading!

Hey Guys,

So what is a Graph Neural Network again? GNNs apply the predictive power of deep learning to rich data structures that depict objects and their relationships as points connected by lines in a graph.

Kumo, a startup offering an AI-powered platform to tackle predictive problems in business, today announced that it raised $18 million in a Series B round led by Sequoia, with participation from A Capital, SV Angel and several angel investors.

Kumo’s platform works specifically with graph neural networks, a class of AI system for processing data that can be represented as a series of graphs. Graphs in this context refer to mathematical constructs made up of vertices (also called nodes) that are connected by edges (or lines). Graphs can be used to model relations and processes in social, IT and even biological systems. In this sense, I believe they will intersect with quantum computing and biotechnology and advanced chemistry.

Quantum Chemistry you say? How so? Most of the current neural network models in quantum chemistry (QC) exclude the molecular symmetry and separate the well-correlated real space (R space) and momenta space (K space) into two individuals, which lack the essential physics in molecular chemistry. In this work, by endorsing the molecular symmetry and elementals of group theory, we propose a comprehendible method to apply symmetry in the graph neural network (SY-GNN), which extends the property-predicting coverage to orbital symmetry for both ground and excited states.

So this is more important than many realize, think about it, today, developers are applying AI’s ability to find patterns to massive graph databases that store information about relationships among data points of all sorts. Together they produce a powerful new tool called graph neural networks.

According to Nvidia, Graph neural networks (GNNs) apply the predictive power of deep learning to rich data structures that depict objects and their relationships as points connected by lines in a graph.

In GNNs, data points are called nodes, which are linked by lines — called edges — with elements expressed mathematically so machine learning algorithms can make useful predictions at the level of nodes, edges or entire graphs.

What Can GNNs Do?

An expanding list of companies is applying GNNs to improve drug discovery, fraud detection and recommendation systems. These applications and many more rely on finding patterns in relationships among data points.

Subscribe now

Researchers are exploring use cases for GNNs in computer graphics, cybersecurity, genomics and materials science. A recent paper reported how GNNs used transportation maps as graphs to improve predictions of arrival time.

GNNs could become more important as A.I. evolves in science and healthcare as well. Lots of learning tasks require dealing with graph data which contains rich relation information among elements. Modeling physics systems, learning molecular fingerprints, predicting protein interface, and classifying diseases demand a model to learn from graph inputs. In other domains such as learning from non-structural data like texts and images, reasoning on extracted structures (like the dependency trees of sentences and the scene graphs of images) is an important research topic which also needs graph reasoning models.

In short, many branches of science and industry already store valuable data in graph databases. With deep learning, they can train predictive models that unearth fresh insights from their graphs.

It turns out I wrote about Graph Neural Networks back in April on this Newsletter, I didn’t even realize. It discovered even some researchers in Quantum computing are looking into it as it relates to the future of simulations and predictions there. The commercial viability is also pretty high, going back to Kumo.AI and what they will specialize in doing for enterprise customers.

Graph neural networks have powerful predictive capabilities. At Pinterest and LinkedIn, they’re used to recommend posts, people and more to hundreds of millions of active users. But at present they are computationally expensive to run — making them cost-prohibitive for most companies.

Graph neural networks (GNNs) are neural models that capture the dependence of graphs via message passing between the nodes of graphs. In recent years, variants of GNNs such as graph convolutional network (GCN), graph attention network (GAT), graph recurrent network (GRN) have demonstrated ground-breaking performances on many deep learning tasks.

Indeed GNNs feel like tapping into some universal language found in nature.

So much of our knowledge in the scientific domain is expressed in graphs, nodes and the relationships between them. Think about it, A.I. in this domain has tremendous potential.

A graph can represent things like social media networks, or molecules. Think of nodes as users, and edges as connections. A social media graph might look like this:

If you think of nodes as customers, and edges as variables in how those customers interact with your product, GNNs have a lot of potential to optimize these relationships.

Kumo remains in the pilot stage but the startup has raised now already $37 million dollars. The predictive power of A.I. is only going to get better once GNNs become cheaper and their utility becomes tailored to different use cases. Clearly this may have some interesting intersections with quantum computing in the future, when those computers are more scaled.

Graph databases have grown from lending themselves to specific use cases such as charting social relationships to supporting a new generation of security and customer-360 solutions. Research firm Gartner has predicted they will drive a trend that will shape data and analytics (D&A) for years to come

We live in a world permeated by relationships of things to other systems.

This reminds me of computational biology and the future of genomics too. Graphs are a kind of data structure which models a set of objects (nodes) and their relationships (edges). Recently, researches on analyzing graphs with machine learning have been receiving more and more attention because of the great expressive power of graphs, i.e. graphs can be used as denotation of a large number of systems across various areas including social science (social networks (Wu et al., 2020), natural science (physical systems (Sanchez et al., 2018; Battaglia et al., 2016) and protein-protein interaction networks (Fout et al., 2017)), knowledge graphs (Hamaguchi et al., 2017) and many other research areas (Khalil et al., 2017).

Recently, some Amazon researchers created a novel method for suggesting related products utilizing directed graphs and graph neural networks. The team started deploying this model in production after the study was presented at this year’s European Conference on Machine Learning (ECML). On comparing model predictions to real customer co-purchases using two performance metrics, HitRate and mean reciprocal rank, the team showed that their method surpassed state-of-the-art baselines by 30% to 160%. They proposed a solution called DAEMON, a Novel Graph Neural Network based Framework for Related Product Recommendation.

I therefore think Kumo.AI could eventually be acquired by someone like Amazon. Amazon Neptune is a managed graph database product published by Amazon.com. It is used as a web service and is part of Amazon Web Services. A lot of A.I-as-a-Service is occurring on the Cloud. First announced at Amazon's re:Invent conference in 2017, Neptune went into general availability in May 2018. The fundamental difficulty with employing graph neural networks (GNNs) for related-product recommendation is that there exist asymmetric relationships between the items.

Read the Amazon Paper

GNNs are “catching fire because of their flexibility to model complex relationships, something traditional neural networks cannot do,” said Jure Leskovec, an associate professor at Stanford (listen above).

While they are used by Amazon in fraud detection and by LinkedIn in social recommendations, clearly they have a lot more utility to be unlocked than that.

“GNNs are general-purpose tools, and every year we discover a bunch of new apps for them,” said Joe Eaton, a distinguished engineer at NVIDIA who is leading a team applying accelerated computing to GNNs. “We haven’t even scratched the surface of what GNNs can do.”

Interesting stuff for the future.

Subscribe now

Hey Guys,

I’ve been hearing yet a new hype-trend, and it’s called “Software 3.0”. But what is it?

As the Cloud and the Modern Data Stack evolves even with GitHub Copilot type AI buddies for software development folk, it makes me wonder. Is software really evolving too? Is Generative-AI or no-code platforms part of the Software 3.0 movement?

One explanation of what Software 3.0 is goes thus: As computing and storage have grown cheaper and digitization has become ubiquitous, enterprises are creating ever-increasing amounts of digital data. The availability of this data, and of cost-effective ways to process it, has led to the emergence of enterprise AI systems, or “Software 3.0”.

But rather, I think we can safely say Generative AI itself heralds the beginning of Software 3.0. I think it resembles all of these newer movements like RPA, no-code platforms, the MDS and the era of Databricks and Snowflake. Software 3.0 also is reshaping how innovation occurs in machine learning.

It’s not all about massive Enterprise trends around data science or consumer apps for machine learning. Another thing that is occurring is AI Labs are parenting with a BigTech parent company. Think Microsoft and OpenAI.

These AI Labs that sometimes start out as non-profits become innovation Labs inside and working side-by-side these massive Enterprise companies.

These compute partnerships and trend for more decentralized organizations bringing academic ideas to fruition is part of the Software 3.0 trend. There’s also far more global cooperation and accessibility to Transformers and LLMs and software learning material than ever before.

Software 3.0 can therefore be a metaphor for the new trends in A.I. and software engineering that relate to our improved technological capabilities as compared to say, around ten years ago. Venture Capitalists also uses Software 3.0 as a kind of new-wave for the newer trends in A.I. I’ve been combing the State of A.I. Report 2022 for insights into what these might be.

In Software 3.0 compute is scaling well.

Think about the Supercomputers for a moment:

Share

Companies like Meta and Tesla seem to think that these machines can give them a competitive advantage in things like the Metaverse or Autonomous vehicles. That remains to be seen.

Software 3.0 in a sense represents a different paradigm:

Software 3.0 businesses create value for customers not by producing software code alone, but composite systems comprising code and data. In this phase, customers and vendors have a symbiotic relationship

Just like Web 3.0 is supposed to be another paradigm, in fact we are likely closer to Software 3.0 just since the rate of technological change is faster while Web 3.0 requires certain institutional, leadership, regulatory and other kinds of changes that are far slower.

Software 3.0 vendors leverage customer data to train their proprietary ML models and tune product experiences, while customers receive added value from the vendor through data network effects arising from the vendor’s customer base. This symbiotic relationship sets up equitable risk- and reward-sharing between the customer and the vendor, but also requires a period of investment from both sides before the full rewards of this model can be reaped.

I would actually suggest the Software 3.0 era begins circa 2024. A nascent period where new kinds of work is being done that more approximates the road to AGI and A.I. more actively involved in software development and more maturity of the low-code and RPA sectors.

It’s nice to talk about the future or hype A.I. or assume software is itself evolving quickly, but the future always takes more time to arrive than we thought. This can be seen with chatbots, VR, self-driving cars and basically any other hype cycle trend like blockchain, quantum computing or human level A.I. (HLAI).

I do think however that it’s useful to define what characteristics and trends Software 3.0 might encompass. If we are to be a data-based society we need to continue to refine our processes of data science, data engineering and MLOps.

The most important way in which Software 3.0 differs from previous phases of evolution in the software industry is its dependency on data. Advanced ML techniques, including deep neural networks, lie at the heart of enterprise AI businesses. These models are trained using sample data points from the relevant problem domain then are deployed into production, where they make decisions for business use.

If we are to say that autonomous vehicles are part of a Software 3.0 native world, then the current broken state of things for self-driving cars doesn’t suggest that we have yet reached the Software 3.0 era. For as good as the MDS is, it still needs a few years to refine itself even as the likes of Snowflake and Databricks continue to mature and all the other related companies, for example.

Venture Capitals enjoying hopping on a band-wagon too. Sarah Guo announced that she raised $101 million for her new fund called “Conviction”, to back companies that are building artificial intelligence and what she describes as software 3.0. I didn’t happen to see any active mention of “Software 3.0” in her blog. I do wish Venture Capitalists had really good content, that made me want to believe in their vision.

You could mine a16z’s Future blog and find some things that resemble Software 3.0, without explicitly defining the new paradigm in this light of datascience, machine learning and software development. Perhaps the hype around Robotic Process Automation comes the closest. I have noticed as well that Robotics is becoming more mainstream in 2022 with more startups and papers working on the problems. Even as FedEx and Amazon have proved that last mile delivery robots aren’t ready for the consumer world yet.

Software 3.0 therefore isn’t just about software, but a world more permeated by A.I. and data engineering. The evolution of MLOps is clearly important in this trend. While a16z’s Future media focuses on less than stellar in terms of realism trends like NFTs, I wish they focused more on core-A.I. hype trends that don’t simply go away anytime there is a crypto winter. So much content about Web 3.0, with so little grounding in reality.

Curiously we do find decentralized orgs have an advantage in the kind of work Software 3.0 era requires.

Meta concluded that “while the centralized nature of the [AI] organization gave us leverage in some areas it also made it a challenge to integrate as deeply as we would hope.”

These new decentralized A.I. organizations are part of the Software 3.0 trend.

These clusters are therefore Software 3.0 native, think about it:

Here is the one for DeepMind (Source: State of AI Report 2022)

We are a new trajectory for more decentralized A.I. Labs that can innovate faster. This is a component of Software 3.0 where we are actually doing something better.

So what is the new breed? It’s still too soon to say, but their ideas and ideology are BIG:

Subscribe now

In Software 3.0 from one perspective, there’s more iterations of the product that’s customer-centric and trust-based. Think of Tesla experimenting with FSD, it isn’t quite there yet but people still pay for it, why is that? Is it just because they like Elon Musk? No they believe that their data will help build the product, even if they over-estimate the level of autonomy Tesla’s FSD itself provides (which is quite dangerous by the way).

The Advent of A.I. Coding Assistants

While also nascent, these coding buddies are also Software 3.0 native, and they imply more productive software development cycles with more peak productive hours for software engineers.

For instance if GitHub Copilot turns out to be a bust, there are other potential winners here: OpenAI’s Codex quickly evolved from research (July 2021) to open commercialization (June 2022) with (Microsoft’s) GitHub Copilot now publicly available for $10/month or $100/year. Amazon followed suit by announcing CodeWhisperer in preview in June 2022. Google revealed that it was using an internal ML-powered code completion tool (so maybe in a few years in a browser IDE?). Meanwhile, with its 1M+ users, tabnine raised $15M, promising accurate multiline code completions.

• tabnine

• Amazon CodeWhisperer

• Microsoft’s GitHub Copilot, etc…

There are actually dozens of Software 3.0 trends already improving themselves, but few reports spot them or outline their efficacy and path to product-market fit. I’m not sure what that is really. Even who the top MLops Platforms are for Enterprise is pretty open to debate. This is why I don’t think in 2022 we’ve entered the Software 3.0 era quiet yet, though there are signs of it out there.

For me another element of Software 3.0 is A.I. showing real ROI in big areas like Pharma and Healthcare. A.I. in drug development is showing a lot more potential, but it takes years to get more street cred in Big Pharma, although there are a lot of interesting startups involved in AI in drug development and it’s a lot about the speed to market. It saves time therefore it saves huge costs, and this is what healthcare needs by and large; it needs huge cost savings as the population gets older in developed countries especially in the 2030s, it’s literally a race against time.

What else might be considered part of the Software 3.0 paradigm?

• A.I. in medical diagnosis

• Prompt based Generative A.I. in text to speech, video, music and writing.

• Task automation for example no-code and robotic process automation solutions

• Increase in industrial and service robots per worker (ratio)

• Smart automation involving IoT, green technology, 5 and 6G, satellite tech

• Embedded communications: like Google’s new tech. Project Starline. (which is legit)

• Evolution of LLMs, Transformers and Supercomputers

• Interoperability advances in Web3

• A.I. spinouts from top Universities (bleeding edge A.I. startups)

Software 3.0 isn’t Inevitable

With higher interest rates however Venture Capital and investments are noticeable slowed.

Private companies using AI are expected to raise 36% less money in 2022* vs. last year, but are still on track to exceed the 2020 level. This is comparable with the investment in all startups & scaleups worldwide.

Anyways guys I hope you have enjoyed my musings on Software 3.0, I will soon return to the more Data science data engineering based topics. But as a future watcher I’m also interested in the macro trends of the 2020s where so much is is apparently “nascent”.

Thanks for reading and if you want to support the channel or the price of a cup of coffee, it goes a long way.

Subscribe now

Hey Guys,

Happy Monday. So it seems like Meta is spinning out PyTorch. Increasingly it seems that The PyTorch Foundation wants to be seen as a “neutral home” for the deep learning community to collaborate on the open source PyTorch framework and ecosystem.

TensorFlow offers better visualization, which allows developers to debug better and track the training process. PyTorch, however, provides only limited visualization. TensorFlow also beats PyTorch in deploying trained models to production, thanks to the TensorFlow Serving framework.

Datascience Learning Center

PyTorch on the Cusp of Disrupting TensorFlow

Hey Guys, PyTorch has really become in 2022, a bit more dominant than Tensorflow. So, both TensorFlow and PyTorch provide useful abstractions to reduce amounts of boilerplate code and speed up model development. The main difference between them is that PyTorch may feel more “pythonic” and has an object-oriented approach while TensorFlow has several opti…

Read more

4 years ago · 2 likes · Michael Spencer

Today an edict came down from the Zuck:

Meta AI believes these are the takeaways:

TL;DR Takeaways

• PyTorch, one of the leading community-driven AI research frameworks, is moving to a new, independent PyTorch Foundation that will be part of the Linux Foundation.

• The PyTorch Foundation will democratize access to state-of-the-art AI tools, libraries and other components to accelerate progress in AI.

• Meta will continue to invest in PyTorch and use it as our primary framework for AI research and production applications at the company.

Sound familiar? Hey Hugging Face and so many others!

What you Need to know?

• To accelerate progress in AI, PyTorch is moving to a new, independent PyTorch Foundation, under the Linux Foundation umbrella.

• The project will join the Linux Foundation with a diverse governing board composed of representatives from AMD, Amazon Web Services, Google Cloud, Meta, Microsoft Azure, and Nvidia, with the intention to expand over time. The PyTorch Foundation will act as a responsible steward for the technology and support PyTorch through conferences, training courses, and other initiatives.

• The foundation’s mission is to drive adoption of AI tooling by fostering and sustaining an ecosystem of open source, vendor-neutral projects with PyTorch. It will democratize state-of-the-art tools, libraries, and other components to make these innovations accessible to everyone.

• The PyTorch Foundation will focus on the business and product marketing of PyTorch and the related ecosystem. The transition will not entail any changes to PyTorch’s code and core project, including its separate technical governance structure.

• Meta will continue to invest in PyTorch and use it as our primary framework for AI research and production applications at the company.

But what’s really going on here?

A Facebook Framework - PyTorch

In 2016, Meta (then but a simple country Facebook) launched its open-source AI research library, the Pytorch framework. Six years and 150,000 projects from 2,400 contributors later it wants to spin in out. And, who can blame them? Facebook and Meta’s reputation is significantly compromised and leading to a employee churn.

I think TensorFlow adoption is out competing PyTorch at this point even in research. How BigTech molds foundations and open-source communities is really interesting to behold. The PyTorch Foundation is supported by leading contributors to the PyTorch open source project. The Foundation leverages resources provided by members and contributors to enable community discussions and collaboration.

Community collaboration is critical for the framework’s evolution as well as the development of associated projects that support using PyTorch in production and at scale. As part of The Linux Foundation, the PyTorch community will also collaborate on training, local and regional events, open source developer tooling, academic research, and guides to help new users and contributors have a productive experience.

It all really makes me wonder.

PyTorch Foundation: A new era for the cutting-edge AI framework

But Pytorch isn’t just Meta’s baby, it serves as a technological underpinning to much of Amazon’s Web Services work as well as Microsoft Azure and OpenAI. It’s known for its easy of use for those who work in Python.

So supposedly Meta cedes control of the open source project in favor of the newly-formed PyTorch Foundation, without "any of the good things" changing. It’s a bit like Google “spinning out” stuff recently.

Meta’s creations are the last places I’d ever call a “neutral home”, so I find the PR on this really deceptive and puzzling.

What do you think?

Leave a comment

It sounds like yet another consortium of BigTech, if I have to be honest with you. , as the Pytorch Foundation, “will boast a wide-ranging governing board composed of representatives from AMD, Amazon Web Services, Google Cloud, Meta, Microsoft Azure, and Nvidia, with the intention to expand further over time.” Oh really? A foundation with a governing board like this? Meta is quite the schemer.

Releases, features, and technical direction will still be driven by the maintainers, committers, and contributors to the project. The creation of the PyTorch Foundation ensures that decisions will be made in a transparent and open manner by a diverse group of members for many years to come.

The hype around the democratization of A.I. well it seems even Meta AI wants a piece of it.

Birth Pangs at FAIR Facebook AI Research

In 2016, a group of Meta AI researchers embarked quietly on a couple of projects. They were determined to create a single, simple, standardized interface for their end-to-end workflow.

They also harbored hopes of fixing the tedious, complicated research-to-production pipeline of the AI field. In fact, this pipeline was more of a labyrinth, involving multiple steps and tools, fragmented processes, and navigation between different frameworks that were optimized for either research or production, but not both.

The Meta team experimented with machine learning (ML) frameworks such as Theano and Torch, and with advanced concepts from Lua Torch, Chainer, and HIPS Autograd, but stayed focused on one thing above all for their new framework: usability.

Just two years later, Meta announced PyTorch 1.0, a dynamic, interactive framework that allowed developers to not only experiment rapidly but also seamlessly transition to graph-based modes for deployment.

Now six years later, they are sharing it with the world!

As reported by Engadget on Yahoo, despite being freed of direct oversight, Meta intends to continue employing Pytorch as its primary AI research platform and financially support it accordingly. Zuckerberg did note however, that the company plans to maintain “a clear separation between the business and technical governance” of the foundation.

Zuck strikes again. Mark Zuckerberg’s poor leadership skills are slowly dragging Meta toward failure, a Harvard expert says. His centralization of power truly has been devastating for democracy and the reputation of silicon Valley in recent years. I’m not sure how many machine learning experts really care though?

When things feel like BigTech Consortiums though, it makes me a bit uncomfortable.

Facebook loves consortiums as I realized in its failed Libra stablecoin attempt.

PyTorch has since grown into the lingua franca of AI research. Today, more than 80 percent of researchers who submit their work at major ML conferences, such as NeurIPS or ICML, harness the framework.

In reality we are entering a deepfake era of the internet, where synthetic AI and Bitcoin’s FOMO is going to distort our reality even further.

Check out this Jeff Bezos and Elon Musk Star TreK Deepfake:

Meta has put open science at the core of our work in AI, whether it’s releasing code for large language models, self-supervised computer vision systems, innovative new datasets, embodied AI platforms, and much more. Meanwhile more of its leadership and top AI talent continues to leave the company.

Meta’s stock is down more than 50% year to date.

TensorFlow is an open source framework developed by Google researchers to run machine learning, deep learning and other statistical and predictive analytics workloads.

Join 33 other paying subscribers for additional content and locked archive posts.

Subscribe now

I’ve started a new Newsletter where I will try to post content for software engineers and technology workers called Data Analytics Future, which might evolve into something interesting.

Credit: the artist, or the A.I. Tool, or I dunno.

Hey Friends,

As you know I a commentator on A.I. ethics and I have my own take on what the Synthetic Internet will do to us.

Allen's winning image looks like a bright, surreal cross between a Renaissance and steampunk painting. (read on)

Meanwhile, in the real world, there has been a stealth win for AI-generated art inspires heated ethics debate on social media. The deepfakes are coming. So what’s all the fuss about?

AI art generators have made a huge impact in digital art in recent months. The latest generator of tools, including DALL-E 2, MidJourney and Stable Diffusion are capable of creating an incredible range of images based on text prompts – you basically tell them what you want them to create, and you are off to the races.

There’s now even a Substack for AI Storytelling:

Stories by AI

Unique tales written, illustrated, and narrated with AI

I’m not really a believer, but our Reddit fueled A.I. commentator seems to think this is a big deal too.

The Algorithmic Bridge

While AI Artists Thrive, Detractors Worry—But Both Miss What's Coming

Last week, game designer Jason Allen used the AI model Midjourney to create “Théâtre D’opéra Spatial,” the artwork you see in the cover image. It’s a nice piece that, although exudes a clear AI scent, has found a delicate mix between traditional baroque opera style, reflected in people’s clothin…

Read more

4 years ago · 6 likes · 5 comments · Alberto Romero

So here’s what happend.

Jason M. Allen was almost too nervous to enter his first art competition. Now, his award-winning image is sparking controversy about whether art can be generated by a computer, and what, exactly, it means to be an artist.

Here is his image:

Jason Allen's AI-generated art won a prize at the Colorado State Fair (Image credit: Jason Allen via Midjourney)

His official statement:

'I’m not going to apologize for it. I won, and I didn’t break any rules.'

Unfortunately many of his peers don’t feel the same way.

A synthetic media artist named Jason Allen entered AI-generated artwork into the Colorado State Fair fine arts competition and announced last week that he won first place in the Digital Arts/Digitally Manipulated Photography category, Vice reported Wednesday based on a viral tweet.

I found this YouTube take on it a bit funny:

I think it’s kind of cool and a sign of the times:

TL;DR

So according to CNN and others: In August, Allen, a game designer who lives in Pueblo West, Colorado, won first place in the emerging artist division's "digital arts/digitally-manipulated photography" category at the Colorado State Fair Fine Arts Competition. His winning image, titled "Théâtre D'opéra Spatial" (French for "Space Opera Theater"), was made with Midjourney — an artificial intelligence system that can produce detailed images when fed written prompts. A $300 prize accompanied his win.

Is this a case of an artist being “augmented” by A.I. or something else? Prompting and creation are definately two different things.

"I'm fascinated by this imagery. I love it. And it think everyone should see it," Allen, 39, told CNN Business in an interview.

Follow this Newsletter on Twitter

No mercy no malice? However, it only emerged after the prize was awarded that Allen had created his artwork using the Discord-based AI art generator Midjourney before upscaling it using AI Gigapixel to put on canvas.

Obviously, this upset some people.

However it’s not a crime against humanity, we just need some ground rules. Unfortunately OpenAI is being disrupted by Stale Diffusion and Midjourney and is not making any ground rules. DALL-E 2 just wants to be better, not exactly share the commercial potential of this kind of technology, with others. OpenAI is sponsored heavily by Microsoft.

Do I actually want to live in a world where GPT-3 helps me write stories? I’m not a fiction writer of an artist, but I can see AI bots doing most of social media marketing that takes place today, it’s not complicated and definately is repetitive.

How do we soul search in a world that has no opt-in switch to potentially spammy A.I. tools? Allen's victory prompted lively discussions on Twitter, Reddit, and the Midjourney Discord server about the nature of art and what it means to be an artist. It’s all rather ridiculous. Even the commentary on this reads to me like glib click-bait. But then again, I’m not looking to trend on Reddit or Hacker News.

Who am I to argue with poetry, I mean music, I mean art that’s not actually very human? It’s still a beautiful piece. Some commenters think human artistry is doomed thanks to AI and that all artists are destined to be replaced by machines. Others think art will evolve and adapt with new technologies that come along, citing synthesizers in music. It's a hot debate that Wired covered in July.

I want to cry, but then again nobody ever listens to my own warnings either.

At least we have NFT profiles able to go viral on Twitter about it, I guess?

Allen, who's the founder and lead developer at the board game company Incarnate Games(opens in new tab), had said that his pieces were created with a computer, thus they were put in the “Digital Art” category. A cunning statement to be sure, and it’s beautiful madness on an unregulated internet where Synthetic AI is actually just getting started. The DOD help found Scale AI that I think is going to be the actual leader in Synthetic A.I. How is that for weaponizing the future?

A.I. can train better on synthetic data and simulations that aren’t real.

PCgamer summarized it as follows:

The arguments for and against this are by now familiar, but boil down to the potential of AI generation software to replace human artistry and skill, and its impact both in terms of costing jobs and the implications for cultural production. There's also the fact that these things are trained on the work of humans who get no credit nor benefit from them.

The author of the Algorithmic Bridge curates the debate a step further. I personally have no idea about this, I’m not about to start arguing on Discord or Reddit about it either. The lack of regulation in the space is just appalling and as A.I. gets more into psychological counseling, education, marketing and sales - there’s going to be some pretty awesome debates, but who will make the profits? It will be BigTech.

Even how this trends on YouTube, Reddit, Hacker News or a Substack article is completely arbitrary and itself fake and click-bait esque. The debates also have zero impact on the outcome.

I’m left feeling a bit nihilistic, in that BigTech itself says A.I. and automation is going to create “all kinds of new jobs”. They omit of course that they will be the major beneficiary from a revenue standpoint of all of these little tools. Who do you suppose will acquire the Grammarlys’ and Jasper.AIs’ of this world in addition to all of this open-source text to image generators?

The weaponization of A.I. occurs each day, and A.I ethics aren’t even being followed by major companies. And the best we can do is be upset about someone hacking an art contest with a deepfake?

Artists (and non-artists) have responded with fury, with many people seeing the use of an AI art generator in an art competition as cheating. Fury on the internet, you don’t say? Is it still trending a few hours later? We’ve created an increasingly fake internet, period.

Unbelievable.

Join 33 other paying subscribers for additional content and locked archive posts.

Subscribe now