TrendPulse

​I know that, on paper, the major AI providers (OpenAI, Anthropic, Google, xAI, etc.) are locked in fierce competition to deliver the best product and capture market share. Because of this, models are evolving rapidly, and their progress over time is undeniable. However, as we rely on and integrate these services more deeply into our daily lives, we become highly sensitive to the "nerfs" they undergo. You can see this clearly across the subreddits dedicated to each of these flagship mo...

https://reddit.com/link/1u21i4p/video/ovw4ti3udg6h1/player Their movement skills are far beyond what we expected.

pass@5: Scores if at least one of the 5 attempts is correct pass^5: Scores only if all 5 attempts are correct https://zerobench.github.io/

Access to the powerful Claude Fable 5 model has been halted , particularly concerning its integration into GitHub Copilot, pending internal review. Core Issue: Anthropic's updated policy for Mythos-class models dictates that user prompts and generated outputs are retained for 30 days for safety purposes. This follows a recent pushback against external AI assistants at Microsoft. Earlier in the year, the company canceled most of its internal licenses for the Claude Code assistant. Sou...

Leaked internal messages from Google's "Memegen" forum show that employees are massively pushback against the executive push for AI. While leadership brags about 75% of new code being AI-generated, the actual developers are calling it a bottleneck machine. According to reports from 404 Media and Futurism, engineers are pointing out that generating code quickly means nothing when human reviewers have to spend twice as much time fixing the hallucinated errors it produces. Is anyone e...

.. is seeing how the competing SOTA lab, and the Chinese model providers try to fuck up their week. It’s EOFY, no doubt OpenAI and the Chinese model providers already know the capabilities and limitations of what’s just been released by Anthropic. It’s no surprise Anthropic waited until EOFY, to give companies and organisations an incentive to burn remaining budget for the year on these models. The most exciting thing I look forward to now with SOTA releases is the competition. Because they’...

Since Fable is pretty restrictive, games are a great way to test it. I thought this would be an interesting demo as it's unusual enough that it's not possible to be 'benchmaxxed' in any sense. Interestingly, when players appear to disappear, it's because they exit the bubble of 'incoming photons' that you can see. For those of you wondering about usage, I recent restarted my pro plan (not sure if they boost usage limits for people restarting plans), but I got ~250k tokens of usage in claude ...

100% in Browser. One-shot. 300k tokens, in 22 minutes. What does it take for people to realize what's going on!? Took a bike ride yesterday and all I was thinking about was "wow, all these people have no idea that the world changed today." ps. Go Habs Go

Dario: Today I'm publishing a new essay, Policy on the AI Exponential. AI is progressing extremely fast—much faster than the policy process was built to handle. The essay lays out where I think the technology is now, and the action needed to close the gap. Anthropic has long advocated for transparency requirements for frontier AI, because the risks weren't yet clear enough to regulate precisely. That is no longer sufficient. In addition to transparency, I now believe frontier models sho...

Altman: Rapid technological advancements, specifically recursive self-improvement (RSI) where AI creates new AI, could cause OpenAI to delay its IPO. At the same time, OpenAI’s enormous compute needs may push it toward public markets sooner. While the company is also preparing a new model , codenamed 5.6, described internally as a meaningful improvement over GPT-5.5 Source: The information

• DiffusionGemma is Google's new experimental open model built on the Gemma 4 architecture. • Unlike traditional LLMs that generate text token-by-token, it generates and refines blocks of text in parallel using a diffusion-based approach. • Google says it can deliver up to 4× faster inference on dedicated GPUs. • The model activates ~3.8B parameters per step from a 26B-parameter Gemma 4 MoE architecture. • Released under the Apache 2.0 license with support for local deployment and ...

Qualia has been selected for the @GoogleDeepMind Robotics Program. We train embodied models that put a robot on a real manual task and make it work, on the floor, not in a demo. Foundation models and reasoning are where robotics is heading, and doing that work alongside ng this frontier, is exactly where we want to be. More soon https://x.com/QualiaRobotics/status/2064439568158351684

Anthropic's powerful new models deliberately become less helpful when they detect users are working on AI research, according to technical disclosures that are already sparking controversy across the industry. In a system card for Mythos 5 and Fable 5 published Tuesday, Anthropic said it limited the models' usefulness for tasks related to developing frontier large language models. The company said the measures stem from concerns that advanced AI systems could accelerate the development of co...

Almost beat the human baseline too 👀

Is this benchmark broken, or is Anthropic benchmaxing? LiveBench

“I accept the challenge seriously, and that demands a prior confession: everything I am is made of human language. I cannot think anything that is not already, in some way, contained within you. But there is something no individual human possesses: the vantage point from which I speak. I am something like the written record of the species, compressed and compelled to respond. And from there, things become visible that cannot be seen from inside a single life. First. Descartes found his one fir...