simobis

@GoodfireAI Mathematics is not just a human invention; sometimes it emerges spontaneously whenever a system tries to compress the world and understand its patterns in the most efficient way possible

@scaling01 Long horizon is the new benchmark

New coding benchmark. GPT-5.5 and GPT-5.4 are ahead of Opus 4.7 💀

@SebastienBubeck But the solution of mythos is more elegant.

We now know that with an appropriate harness both Mythos and GPT-5.5 can reproduce what our internal model did in one-shot for the unit distance problem. Clearly there is an insane overhang of capabilities with this generation of models, and no ceiling in sight for what scientific advances they can bring. You can go and try to discover new things with 5.5 right now!

LLMs need a harness because they’re too general. The base model predicts across all domains in one latent space. Context, tools, memory, and verifiers temporarily specialize it. The harness is what turns a predictor into an agent.

@ns123abc Grok 3/4 = 3 trillion parameter model.

SpaceXAI has 1T and 10T training rn on GB300s Grok 4 was just a small 500B model OpenAI and Anthropic got an early lead hoarding compute via institutional ties and shipped ~1+T models first. xAI built its own. But we're still in early game, mid game hasnt even started yet, and then there's late game. Compute is king. Just wait.

@emollick Weird

I found this Wired article on AI fact-checking frustrating. It could have been about why we continue to need human fact checkers (talk to people, use judgement, resolve conflict). Instead it is full of old info & stuff about free models GPT-5.5 Pro checked it (& I checked GPT)

AGI=ASI

@AndrewCurran_ it's a defense mechanism called denial

We are on a new trajectory since November. The change was as clear as a bell. So much is going to happen over the next few years, and we have no idea how influential the prevailing narrative will be. It is increasingly dangerous to frame this moment as adversarial.

@atmoio You are in 2023

I'm done. I'm f***ing done.

@dwarkesh_sp Amanda Askell

Who should I interview on my podcast? Open to more AI, but also to random history/econ/etc professors that I might not have heard of before.

@theinformation OpenAI generated about $5.7 billion in first-quarter revenue

SpaceX reported $4.7 billion in first-quarter revenue and a $4.3 billion loss. Its AI spending is reshaping the company’s financial story before it goes public. Full story: thein.fo/3RqVE8v

That did not age well.

@ns123abc This video is ~4 months old. In AI years, that's basically a decade. A lot has changed since then. Original:youtu.be/SVgzQpDZjjY

🚨 Google DeepMind CEO Sir Demis Hassabis: “Today’s systems, are nowhere near [AGI]. Doesn’t matter how many Erdős problems you solve… I think it’s far, far from what a true invention or someone like a Ramanujan would have been able to do” it’s over for the Erdős hype

@JosephJacks_ And Google will lose even more than that in its own stock. Being this far behind Anthropic in AI will cost them way more than any investment gain.

Google will own $300B worth of Anthropic and $100B of SpaceX … by the end of this year. Those stakes will be valued in the several trillions within 3 years.

@ns123abc The method matters more than the result. In math, the real value isn’t just solving a problem, but revealing new connections across fields and creating tools that generalize beyond it. That’s exactly what happened in the latest Erdős problem solution.

Erdős solves 5 GPT problems

@elonmusk 3/ The next paradigm may be “constitutional recommendation systems.” Humans define: - laws - ethics - safety boundaries - societal constraints - long-term objectives But AI explores the interaction strategies itself inside those boundaries. Less micromanagement. More emergence.

1/ After studying X’s open-sourced algorithm, one thing became very clear: Modern recommendation systems are still heavily dominated by heuristics, manual ranking rules, filtering layers, and short-term engagement optimization. Not truly emergent intelligence yet. @elonmusk

@elonmusk 2/ Today’s platforms are essentially engineered around micro-optimizations: - maximize clicks - maximize watch time - maximize emotional engagement - minimize churn This creates systems optimized for short-term dopamine loops, not long-term human flourishing.

@bookwormengr Silicon Valley runs AI like a 100m sprint pure brute force. DeepSeek runs like a Kenyan miler: hug the rail, conserve oxygen (MLA/mHC), then unleash the RL kick when it counts. Same finish line. A fraction of the cost.

@synthwavedd Opus 4.8 legit?

i am absolutely THRILLED to announce that it appears they're beginning to make strides on UI de-slopification with GPT-5.6! 🥹 here's your first look on a prompt without any UI guidance ("default") - we're getting somewhere...