Next year, we will have GPT 6.2 level intelligence at 1,000 tokens per second. We are also going to see the mass manufacturing of multiple humanoid companies in America. Tesla and Figure have already completed their mass manufacturing pilot lines.

New Opus/Fable checkpoint animation I asked it to make an animation in the mobile app about a girl and a robot in the forest who are chased by a skinwalker. No assets, no references, or anything like that. I was debating on uploading it because the quality wasn't up to my standards, but for a model with no references, no Claude Code harness, or anything like that, just in the mobile app, this is a pretty good comparison to what Opus 5.5 was able to do.

Anti-AI and Anti-data center hysteria is literally just a repeat of the Anti-nuclear hysteria of the 70s. That one cost us decades of advancements in the field of power generation and the cheap, abundant energy it would have brought. Don't repeat it with this new tech.

I urge you to read this post. Let's play it out. Let's pause American AI. Let's ban RSI. Put a moratorium on all the data centers. America's progress will severely slow down. China's will inevitably keep going. They don't care about your American moratorium. They will continue to build RSI anyway, (this is guaranteed) and they will most likely keep open-sourcing models like they already do. Now what? In two to three years' time, you're going to have every American company and every small business secretly downloading open-source AGI. And if you say you're going to ban Fortune 500 companies from using Chinese models, you still can't stop the weights from spreading. What are you going to do, ban math? So the job loss happens anyway. I have my own small business, and I also work for a major company. The side teams are going to use it like they already do, I'm going to use it, and people in other small businesses are going to use it to replace editors, gig workers, programmers, and software engineers who might have done small projects here and there. Even with the YouTube teams I work with, they already plan on replacing all of their editors in one years time. - they will just simply switch to Chinese models. The job loss is going to happen regardless: the editing, the assistants, customer service, and a huge chunk of gig work will all be done by free open-source models from China. And unless you're banning self-driving cars too, the majority of rideshare will eventually go with it. Except now, the upside is no longer in America. The labs, the chips, and the tax revenue are all going to go to China. And with the data center moratorium, even running those models happens on someone else's servers. We get the unemployment and no growth to pay for UBI. You're going to completely fuck over every single American, whether they like it or not. If you think advanced AI is dangerous, you still have to explain how an American pause protects us from systems developed abroad and whether it leaves us less able to defend ourselves. I've spoken to a plethora of Chinese researchers at a myriad of companies, gathering my consensus, and I am confident that they will not stop development. No matter how America tries to come to the table, they will continue pressing forward to AGI. We've had hundreds of cyberattacks from Chinese state hackers that have exposed critical vulnerabilities. Not only are you going to ramp that up, but you're going to do it in tandem with looming job loss.

It's unfortunate how no-name, lackey, layman, insipid congressmen have to resort to saying terrible things about AI just to become popular. These people have no understanding of alignment research and want to automatically ban things that sounds scary to them. Story of a politician, I guess.

Remember when OpenAI shared their revenue targets and everyone thought they were insane and that they would go bankrupt at any moment. Now they’re more than 2 times ahead of their 2026 projection.

Anthropic found Claude Sonnet 3.7 getting stuck managing its own memory while playing Pokémon. “Over the next 17 steps, it created six files, deleted nine others, and took no actions in the game itself.” In one run, it wrote 413 files and earned one badge. A stronger model kept just nine files and earned four badges in its run.

How much better is each internal model getting at Anthropic? Model 2: - 174.1 Model 3: - 177.0 (+2.9 points) Model 4: - 178.6 (+1.6 points) That’s a 4.5-point increase from Model 2 to Model 4. Model 2 and 3 are estimates read from METR’s chart. Keep in mind this is AECI, not Epoch’s ECI They use different methods, so we can’t tell from this whether their ECI scores would be higher. Even if you assume Model 4 is Fable 5.5, the changes are about 1.7% and 0.9% of the index scores, but AECI uses an arbitrary scale. For reference Astra scores 169 on the ECI

🚨Kimi k3.1 (Kibitz on Arena) will release before Tuesday next week. It will be a state of the art Open Source model at low cost. “build what you think you look like in Minecraft” GPT6 ASTRA 5,738 BLOCKS • $5.31 KIMI K3.1 13,707 BLOCKS • $0.38

Last week, I teased a significant healthcare discovery at Anthropic, specifically mentioning Mythos 5 and its role in the scientific discovery process. Well, Anthropic just published the research! Working alongside Citrini Research, an internal Claude model autonomously generated over 318,000 protein designs in just 64 hours. Of the 211 candidates selected for laboratory testing, 17 successfully bound to four dangerous scorpion neurotoxins. Keep in mind, this project was initiated by a financial investment research firm, not a biotech company. Non-experts were able to direct an AI system through a genuine molecular discovery campaign and produce experimentally validated results. The implications here are enormous. We are quite literally democratizing access to biological research in ways that could accelerate the discovery of treatments for rare and neglected diseases, potentially helping millions of people who have historically been underserved by traditional drug development.