Chris

A second Jensen Huang affair allegation just hit the timeline, and there’s reportedly video 🥲 The New York Times is reporting that Jensen Huang allegedly had a relationship with another Nvidia executive, Kimberly Powell, who reports directly to him. This is separate from the previously reported relationship with Alice Huang, the Chinese businesswoman linked to an espionage investigation. The report also describes an alleged video of Huang and Powell being affectionate. The Times says it has not seen the video.

Nvidia is SOTA on affair bench

Finally an agent first smart phone.

Strauss Zelnick specifically made it very clear rockstar & its sister companies have used generative AI specifically for creating new mobile game levels. However after that statement he specifically mentioned rockstars cost profile - and said “the cost savings have been small” GTA 6 is part of the cost profile. Perhaps they could be doing something as simple as bug fixes with Claude Code. There is no way on that tight deadline they weren’t using Claude code or codex during the last months of development in my opinion

OpenAI’s unreleased internal model is sitting above Claude Fable 5.2 on the BenchCAD custom-harness leaderboard OpenAI: 87.5% Claude Fable 5.2: 83.5% However OpenAI’s result used the older 2.0 scorer. BenchCAD says scores across harnesses aren’t directly comparable despite them still listing, OpenAIs score underneath BenchCad 2.1

I was watching one of my favorite CEOs' interviews regarding GTA 6, and I couldn't help but think at this specific clip: despite how much Rockstar talks about how they don't use generative AI and how they want to keep humans in the loop,- that he sort of implied GTA 6 was using generative AI here. Strauss Zelnick. was discussing Take-Two’s costs broadly and didn’t confirm whether GTA 6 used it. The reason I wonder if it could be included is that he later gave more specific examples of AI use in mobile games, as well as savings on marketing for some of Take-Two’s studios. Earlier in his answer, he described GTA as handcrafted. About a minute later, discussing AI’s benefits and Take-Two’s overall costs, he said: “I haven't really seen much of that yet. And in the context of our cost profile, the cost savings have been relatively small.” If you're talking about the cost profile of the company itself, it's very unlikely that this wouldn't branch into some of GTA 6's development, especially with how good the models are today. Perhaps this is just the first domino that falls for Rockstar in terms of admitting their generative AI use. I do believe it is inevitable, and they will fall dramatically behind if they don't end up using generative AI, but that's obvious to everyone who knows how quickly this field is moving.

Models are still jagged. We could very well be in a scenario where we have superhuman R&D, but models still lack in common sense. One interesting thing about SimpleBench, (my favorite benchmark), is that we've recently seen a noticeable deviation from its otherwise remarkably consistent trend line. The newest models are outperforming what that trajectory would have predicted, the good news is it looks like they are operating on a different substrate of capabilities. We've seen only linear, gradual progress on SimpleBench despite exponential gains in other domains, but most people don't understand that common sense, intuition, and word logic take a much longer time to improve over domains like mathematics and coding. We might be in a scenario where models are able to be heavily involved in their own R&D and might even have almost full RSI capabilities, but still lack in some common sense areas that would be easy to humans, which is fascinating to me.

Qwen 4 Max is coming!!

Imagine all the forecasters and everyone’s wrong, and we just end up hitting superhuman RSI by the end of the year.

Incredible that models of this capability still can’t beat humans at word logic puzzles.

If progress continues to hold. Not accelerate, the AI2027 forecasters are predicting a superhuman AI researcher by 2027. Previously 2030.

Hedra is giving teams a way to train models on their own video and sensor data for the physical world. That means the data coming from your cameras and devices can be used to build models for your specific environment instead of relying entirely on a generic model.

I’ve heard Anthropic’s model 4 is “very different” in terms of cyber and RSI capabilities as compared to model 3

Specifically the number they give for their current-pace extrapolation is a 65% chance of a superhuman AI researcher by the end of 2027. That’s without assuming progress speeds up any further. The recent pace just has to hold.

This 15 person team is out preforming alpha fold! Geodesic’s NovaAtom-Lite predicts proteins, DNA, RNA and small molecules together. The team reports 79.07% success on FoldBench’s antibody–antigen task vs 47.90% for AlphaFold 3, and 77.06% vs 64.90% on protein–ligand. Capital is already moving for AI for science: Isomorphic ~$40B, US gov + Google + Meta + Biohub $1.8B, first AI-designed drug in Phase III.

Agents shouldn’t need the entire company history pasted into every prompt. Graphon is building a persistent graph to track how information connects and changes over time. Give an agent the relationships and sequence, not just a pile of retrieved documents.

They laughed at me when I said Anthropic will saturate the original SWE verified bench by mid 2026, btw. They told me “extrapolating linearly was unfeasible”

Audio engineers when they realize it made the audio without ears. it has never heard a sound in its life the model is literally deaf 💀