### AI Innovation Challenges: Slowing Progress and New Developments

AI companies are facing challenges in achieving significant performance leaps in their next-generation models.

### AI Innovations and Developments: A Comprehensive Update
* 📰 The pace of innovation at major AI companies like OpenAI, Google, and Anthropic appears to be **slowing**, as they struggle to deliver the **next generation of models**, particularly with OpenAI's Orion model **not meeting performance expectations**.
* 📉 While earlier iterations like GPT-3 showcased significant advancements, the subsequent models have not consistently achieved the **same level of improvement**, similar to how newer iPhones often feel **less revolutionary** than their predecessors.
* 🔍 Google's Gemini software and Anthropic's Claude 3.5 model are also reportedly underperforming against internal benchmarks, highlighting **industry-wide challenges in achieving substantial performance leaps**.
* 💡 Despite these setbacks, companies are expected to continue innovating, as **substantial investments in AI** signal that **breakthroughs are likely just around the corner**.
* 💼 Greg Brockman has returned to OpenAI after an extended leave, indicating a **shift in leadership dynamics** as he seeks a **technical role alongside Sam Altman**.
* 🌐 The domain name **chat.com** was purchased by OpenAI for **over $10 million**, showcasing strategic branding efforts, while Ollama's new version 0.4 introduces **vision capabilities**, enhancing its functionality for users.
* 🤝 In a notable partnership, Palantir is set to integrate Claude models into **secure environments for U.S. government operations**, raising questions about the **implications of AI in sensitive sectors**.
* 🔧 Meanwhile, advancements like the **Tiny Troop project** from Microsoft and the release of **AlphaFold 3** signal ongoing efforts to **push the boundaries of AI technology** across various applications.

AI news is back! We have so much to talk about today. In our first story, it seems like the pace of innovation at AI companies is slowing down according to this article by Bloomberg. OpenAI, Google, and Anthropic are struggling to build their next generation of models, so...

let's start with ChatGPT. In this article, it says the model known as Orion did not hit the company's desired performance, according to two people familiar with the matter. Now, when we first saw ChatGPT personally, and I know a lot of people were just absolutely...

blown away by it, and then we saw GPT-4, which was a huge improvement. But similar to almost every other technology, each iteration is not going to have that same level of improvement every single time. This is why when you get a new iPhone, it...

doesn't feel all that different than the previous two or three generations. Now, in my mind, the GPT-3 models were a huge jump in performance, and it also allows for a lot of headroom in terms of scalability. But a lot of these companies are struggling to really hit these big...

performance jumps of their next-gen models. According to the article again, at Alphabet, Google's upcoming iteration of its Gemini software is not living up to internal expectations, according to three people with knowledge of the matter. And Anthropic has seen the timetable slip for the release of its...

long-awaited cloud model called Claude 3.5. Opus. Now, we already know a lot of the bottlenecks that these companies might be facing, whether it's compute or available data. But, just based on some of the potential post-training unlocks that are available, whether it's...

Chain of Thought or test-time compute, and now I'm about to publish a video about test-time training. There's a lot of performance and scaling availability just based on algorithmic unlocks, but this news story seems to be everywhere now. I don't think these are huge

stumbles for these companies. They are going to continue to innovate. There is so much money being poured into AI in general, so it's just a matter of time until they really start to leverage some of these new techniques. Next, Greg Brockman is back. There was some doubt

that Greg Brockman would come back at all. He was on an extended leave, on vacation basically, his first vacation in many years since joining OpenAI. Now he's back, and him and Sam Altman are working together to find the right role for him. He wants to be in a technical

role, which is great, and I'm glad to see at least somebody is coming back to OpenAI. And our next story, chat.gpt.com is now just chat.com. It is so much cleaner, as you see here, drop the GPT, it's cleaner now. There is actually a

funny story behind this domain name. First of all, it definitely wasn't cheap, and here's the story. So Dharmesh Shah, the founder of HubSpot, said two months ago, and this was about a year ago, two months ago, I bought chat.com for $10 million

plus. Here's an update. I'm posting this because several folks have noticed that the chat.com domain no longer redirects to my post, and some are speculating that I might be getting ready to launch something big on it. Want to clear up confusion. I'm not launching anything. I

sold the domain, so he sold it to OpenAI. Now we know this arbitrage right here probably earned him a few million dollars. Incredibly impressive for basically no work at all except, of course, having about $10 million in capital to invest in the domain to begin.

With that, congrats to Daresh. Great win. Congrats to ChatGPT, chat.com, great domain name. Next, Ollama released a new version 0.4 which now has vision capabilities, specifically support for Meta's Llama 3.2 Vision, both the 11

billion and 90 billion parameter models. So here's a few examples of what it can do. You can feed it your doctor's prescription, which is notoriously hard to read, and it just outputs it right there, no problem. Here you can upload a receipt and ask questions to it. Here you

can get a graph and ask questions to it. So I've tested the Llama 3.2 Vision models. I did not have great results from it, but I think I'm going to try it again, and I have a feeling it's going to do

better. There's probably something I was not doing right. But either way, this is a great update. Congrats to Ollama. Next, Palantir and Anthropic are partnering, and Palantir is bringing the Claude series of models to private secure environments. Palantir will be the first industry partner

to bring Claude models to classified environments for US government intelligence and defense operations. And potentially the scariest thing you can see as a model provider or a partner on AI is Palantir, the Liberator, giving you the eyeballs because that means it's only a

matter of time until he jailbreaks your system so what do you think about this Palantir has been exploding in their stock price lately and they really seem to be making a lot of important moves now bringing this kind of level of artificial intelligence to the

government seems like a mixed bag I want to know what you think drop your thoughts in the comments next Quin 2.5 coder is now out this is potentially the best coding model open source at least out there and these are all available on

Llama now so 32B 14B 7B 3B 1.5B all the way down to 500 million parameter model and Quen launched it together with Llama so these are great coding models and we can see its performance right here on

Live Codebench and MVP3 shot compared to other comparable models next Microsoft dropped a new open source project called Tiny Troop and I actually posted about it and a lot of people liked it it got 230,000 views on this so it was the top

trending model in GitHub and Tiny Troop is an experimental Python library that allows the simulation of people with specific personalities interests and goals kind of like Tiny World basically you're setting up all these agents to just interact with each other you give them personalities and it's really

interesting it's an interesting project I always like seeing projects like this there have been a number of them but this is by Microsoft and I want to test it out if you want to see a tutorial if you want to see me test it in a full

Video, let me know in the comments next. TSMC has been forbidden to manufacture 2 nanometer chips outside of Taiwan, which is kind of nuts to think about. And what they're thinking is they need to protect their intellectual property, they need to protect these chips. And I just reported

a couple weeks ago that TSMC is manufacturing chips in the US. The yields are incredible already, and now this might hinder that initiative. TSMC won't be allowed to produce cutting-edge nodes outside Taiwan, hinting towards a fallout in US-Taiwan relations. And it looks like

Taiwan isn't happy with the progress TSMC is making in the US, especially after remarks by president-elect Donald Trump who claimed Taiwan is stealing US semiconductor technology. So now we're going to have chip wars it seems. Next, AlphaFold 3 is now out. AlphaFold 3 as

a reminder, AlphaFold predicts the folding patterns of proteins to allow us to predict proteins and their behaviors. And a number of people from the DeepMind team actually won Nobel prizes for their work on AlphaFold. And now AlphaFold 3 is out, it is open source, you can

play around with it yourself. So here's the repository, I'll drop it in the description below. And next, it seems like every week we are getting a new diffusion-based video game engine, which means no actual code, no actual video game engine, it is just predicting the

next frame. And that's what we're seeing here, Lucid V1, which surprisingly is open source. So it is described as a world model that can emulate Minecraft environments in real time on consumer hardware. So here are a couple examples. Okay, so this looks like Minecraft, but

just remember this is playable and completely prediction based, completely diffusion based. There is no deterministic game engine here. It is so cool to think about, this is absolutely the future of video games. So you can even play around with it in this Verse

app. I'll drop a link to this in the description below. Let's give it a try real quick and there it is, look at that. It's obviously far from great but it works and again, completely diffusion based, completely AI based end-to-end. If you want to play around with it,

please do. Next, Waymo continues to expand. They are on a tear now. 247 trips are available in Los Angeles, which is where I'm from, now where I live, a little bit outside Los Angeles. You don't really see them anywhere, but I just went to the

city a couple days ago and I saw them everywhere. It's so cool to see. I actually haven't ridden one yet and I really want to. So now the robot taxi is available to anybody in Los Angeles and previously you can only ride if you were part of

their invitation program, but now they are letting everybody in. 80 square miles of LA County is available. Next, we now know a little bit more about what Mir Moradi, the former CTO of OpenAI, is up to. Manana Chen, a research program manager at OpenAI responsible for

preparing AI models for release, has recently left to join Mir Moradi at her new venture. She is also teaming up with ex-head of post-training, Barrett Zoph, and former senior researcher, Luke Metz, both of whom we've reported

have left the model developer OpenAI in recent months. Chen's loss is particularly significant given that she led the launches of a number of OpenAI models and products including GPT-4, its 01 reasoning models, and advanced voice models. Three absolutely incredible launches. It's still not clear what Mir

Marat's company is actually going to be doing, but I'm sure we'll find out soon. Next, we have a new incredible demo video of a robot out of China. Look at this. So, thank you to Chubby for posting this. I think it's incredible. And by the way, if

you're not subscribed to the Forward Future newsletter, you're missing out on some incredibly insightful deep dives into AI by Chubby himself. He has been writing an absolutely incredible series and a number of other one-off articles that are fascinating and you learn a ton. And yeah, so Forward Future, definitely go

subscribe, check it out. But look at the agility of this robot. Not only can it climb and move its legs, but obviously, it can also roll and climb up rocks and climb down rocks and be on two legs or four legs. I mean, this is absolutely

insane to watch. This is by a company called Deep Robotics, and they also have a humanoid version, but obviously, the dog version is very, very cool. Next, Naous Research is releasing a new feature called Forge reasoning API beta, and it is described as an advancement in

inference time scaling that can be applied to any model or set of models. So basically, inference time scaling, test-time compute, the same thing that 01 is basically doing except now research allows you to apply it to any model that you can upload to them. The Forge

reasoning engine is capable of dramatically improving Hermes 7B to reach parity in some categories with OpenAI's GPT-3 full at the cost of more inference compute, so basically being able to give open source models the ability to do that test time compute

which we know improves the quality of the output greatly, so they've developed three architectures that this is based on: Monte Carlo tree search, Chain of Thought, and Mixture of Agents, all of which again can be applied to open source models. This all sounds like what we've been reading about Q-learning for a

while, and it ended up being a one-off, and now we have this, which we can apply to any open source model. Next, the Meta AI Ray-Ban glasses are getting some major updates, so from David Woodland, the product lead for the Meta AI Ray-Ban

glasses, sending videos using voice share videos hands-free, share Meta AI responses and Facebook, Instagram contacts with voice text actions and visual reminders, improvements, incoming call announcements, and Meta AI chatbot features are now available in the UK, so finally the EU is getting some more AI love. So really cool

updates. I still love my Meta AI Ray-Ban glasses and I can't wait to try some of these features out. Next, Amazon is getting set to release their own AI chips to rival Nvidia. So not only can you provision Nvidia chips on AWS, but it

seems like you're going to be able to provision Amazon chips on AWS. The new chips are set to release in December according to this article, and it is all to reduce reliance on Nvidia. The upcoming Trainium 2 AI chips are designed

for training large models and are already being tested by companies like Anthropic, Databricks, and Deutsche Telekom according to AWS's VP of Computer Networking Services. We want to be absolutely the best place to run Nvidia, but at the same time we think it's healthy to have an alternative. Yes, 

platform risk definitely. And this is a great way to diversify their portfolio. So that's it for today. If you enjoyed this video, please consider giving it a like and subscribe and I'll see you in the next one.

AI Progress Limit, Chat.com for $15m, Qwen Coder, Ollama Vision, Ex-OpenAI CTO Plans

Artificial Intelligence (AI), Open Source, Generative Art, AI Art, Futurism, ChatGPT, Large Language Models (LLM), Machine Learning, Technology, Coding, Tutorials, AI News, and more

** Exclusive Pinecone Partner **

Matthew Berman

Bloomberg article reports slowing innovation at OpenAI, Google, and Anthropic.

OpenAI's Orion model underperformed expectations.

Incremental improvements in subsequent AI models are becoming less significant.

Google's Gemini and Anthropic's Claude 3.5 are facing development delays.

Bottlenecks include compute resources and data availability.

OpenAI Personnel Changes and Domain Acquisition

Greg Brockman's return to OpenAI after an extended leave.

Brockman and Altman seeking a technical role for him.

OpenAI purchased the chat.com domain for a significant sum.

Ollama 0.4 released with vision capabilities supporting Meta's Llama 3.2 Vision.

Palantir and Anthropic partnership to bring Claude models to secure environments.

Quin 2.5 coder released, an open-source coding model.

Microsoft releases Tiny Troop, an open-source Python library for simulating agents.

Geopolitical and Industrial Developments in AI

TSMC restricted from manufacturing 2nm chips outside Taiwan.

New diffusion-based video game engine, Lucid V1, released as open source.

Autonomous Vehicle and Company Developments

Waymo expands autonomous vehicle services in Los Angeles.

Mir Moradi's new venture recruits former OpenAI employees.

Impressive demo video of an agile robot from Deep Robotics.

Naous Research releases Forge reasoning API beta.

Meta AI Ray-Ban glasses receive major updates.

Amazon prepares to release its own AI chips to compete with Nvidia.