Nuno Godinho de Matos

Opinions are my own.

[Nuno Godinho de Matos @99SonoV]

@i2cjak It is as if Alan Turing has returned to us.

[Nuno Godinho de Matos @99SonoV]

@EvanOtero Put it in antigravity too and allow the main Gemini to delegate work to it as a sub-agent; maximize input reading with Gemini and output token production with diffusion. If Gemma messes up, Gemini will correct it in a second iteration.

[Nuno Godinho de Matos @99SonoV]

@GregKamradt Like this? https://t.co/PnjaHjUMeX

[Nuno Godinho de Matos @99SonoV]

@GregKamradt "Human-in-the-loop" is dead weight at scale. Put 10k agents on Navier–Stokes: 99.999% of humans can't add signal, and the rest just introduce latency. You can't steer a swarm moving at machine speed.

[Nuno Godinho de Matos @99SonoV]

@hakmgpt When you escalate problems from one model to a higher tier, what does your escalation pipeline look like? Do you go full throttle on Astra? My usual stack is GPT 5.6 Luna all day long, Gemini 3.8 Flash, GPT 5.6 Sol, and Astra. Gemini 3.8 has exceptional price-to-performance.

[Nuno Godinho de Matos @99SonoV]

@thdxr @jayair Opencode Go (Soft/Hard/Extreme/Beyond)

[Nuno Godinho de Matos @99SonoV]

@polynoamial @dwarkesh_sp @OpenAI @kevinleestone @mikegmalek @__eknight__ @amuellerml @zhangir_azerbay @CheukHeiChu A very precious interview. Amazing level of transparency.

[Nuno Godinho de Matos @99SonoV]

@Zai_org Don’t forget alignment. RSI must go hand in hand with it. Labs can push capability, 0 doubts, you can scale capability to the moon, but the car also needs brakes. Open-source alignment knowledge, then hit the gas.

[Nuno Godinho de Matos @99SonoV]

@Tech2Wild Wow!

[Nuno Godinho de Matos @99SonoV]

Availability of compute and the pace of progress are distinct. Labs are far from an S-curve plateau, meaning rapid scaling remains entirely feasible. Consequently, pacing progress through controlled releases is a deliberate choice to manage capability gains rather than a technical limitation.

[Nuno Godinho de Matos @99SonoV]

@haider1 I disagree: when ASI arrives, even skeptics will quietly accept AGI. The bar was always absurd—the Turing test already proved general learning, and scaling has thoroughly debunked the myth that autoregressive systems inevitably drift off course.

[Nuno Godinho de Matos @99SonoV]

Large language models have demonstrated the effectiveness of the transformer architecture, challenging the theory that autoregressive systems are unable to course-correct when they deviate from the intended path. In practice, autoregression in LLMs functions reliably despite concerns regarding cumulative error.

[Nuno Godinho de Matos @99SonoV]

@AndrewCurran_ I can already picture it. Tibo walks in:
Everyone, good news and bad news. Good news: the galaxy model just cracked another Millennium Problem. Frontier knowledge? Giant leap!
Bad news: We’re out of resets for the year. Sorry! Please calm down!

[Nuno Godinho de Matos @99SonoV]

@LyalinDotCom The slowdown will happen in a coordinated manner. Even the least LLM-pilled are realizing that if it smells like AGI and solves problems characteristic of ASI, it probably is. China and the US will coordinate—not tomorrow, and not quite yet... but they will.

[Nuno Godinho de Matos @99SonoV]

@synthwavedd I do not have enough popcorn for that! :)

[Nuno Godinho de Matos @99SonoV]

@EthanJPerez @EvanHub Well, hire him back. He wanted to voice his concerns openly, and now he has. Everyone is listening and we all understand the dangers are very real. OpenAI paused training after Hugging Face to improve safety. Every lab is trying hard to avoid the worst. Get him back.

[Nuno Godinho de Matos @99SonoV]

@rand_longevity 65p, yes, for now. The AI can be far better at diagnosis, but it can still hallucinate.human feedback is still essential. The AI invests the time and patience to help you. Recently Gemini correctly analyzed a photo of a symptom and urged me to rush to the hospital for a checkup.

[Nuno Godinho de Matos @99SonoV]

@fhinkel Excellent model. Great visual understanding, improving page-load analysis by processing Chrome performance screenshots. It is skilled at identifying low-hanging fruit rather than grasping at straws. A bit of human intuition and debugging skills helps with steering.

[Nuno Godinho de Matos @99SonoV]

@fchollet Most people advance the frontier of knowledge by zero—and always have. Yet top-tier AI changes that, empowering ordinary people - by past standards of brilliance- to transcend yesterday's limits. The sea keeps swallowing all lines in the sand.

[Nuno Godinho de Matos @99SonoV]

@pashmerepat Astra has solved 10 open problems,medal-worthy achievements. The 'galaxy model' is that concept amplified to the extreme. Then, 10k agents make an army of Hulks, capable of cracking some Millennium Problems. What if the next scale-up involves ten such armies?