Sergio Charles

thesislabs.ai is live

@thdxr @GaryMarcus @Starcloud_ I believe you have a good solution :)

you're probably underestimating how crazy things are

@OpenAI, @thesis_labs will give you your answer in a few days!

@thesis_labs makes the first trajectory more feasible

@Brad08414464 @physical_int

I wonder if we’ll ever see a robotics company that takes off as fast as OpenAI and Anthropic did surely there will be a startup that challenges Tesla the same way OpenAI challenged Google

@fchollet Robots

Current AI is a librarian of existing knowledge. Science requires an explorer of the unknown. You don't win a Nobel Prize by staying in the library.

@paulg Robotics is what takes us to a future of abundance. AGI in a bottle alone is not sufficient. The notion of value also becomes tenuous in such a timeline

"Anything made before 2028 is going to be valuable." — an OpenAI employee implicitly discloses their timetable

@tszzl @rhydhimma How so? AF seems quite revolutionary. Not trying to downplay the raw power of coding agents, as they could in theory help develop an AF, but what AF represented at the time was decades of effort like coding agents

@rhydhimma they are far more significant than alphafold and it’s not close imo

Codex and Claude Code are probably the most revolutionary products of this century. For now. Maybe not as significant as Alphafold, and all the PhD who slogged to get protein structure data. Data is the keyword.

@itsandrewgao @willdepue @thesis_labs will give you a good fix for this shortly! No need to worry about compute woes

nooo how am i supposed to parameter golf when there are no 8xh100s help @willdepue

@elonmusk @demishassabis Who's running the simulation and where are the aliens?

@demishassabis 𝑝(simulation)≈1 However, within the simulation, hardware is extremely hard to do. Only those who have bled on a production line can understand. 

Matter, Energy & Intelligence

@Yuchenj_UW @mimighost008 You must read it after you ipo

@mimighost008 that guy? I haven’t read or watched Harry Potter.

“Dobby runs my entire house over WhatsApp.” Andrej Karpathy is running OpenClaw like Tony Stark runs JARVIS.

One day soon, @thesis_labs will answer all of these!

Absolutely. At @thesis_labs, we're accelerating ai r&d by delegating the tasks you don't want to do to a swarm of research assistant agents. As per Jevon's paradox, this should just accelerate the rate at which you do research itself. Think things like tinkering with hparams, writing experiment code for new hypotheses, configuring multi-gpu runs etc.

Most science isn't hill climbing. ML research that people *want* to do also isn't usually hill climbing. Publish or perish means a lot of ML researchers are (regretfully) doing hill climbing If we can delegate that to agents, ML/AI research itself should flourish Research taste/high level strategy will presumably also get gobbled up by AI, but that (post)training loop is *way* harder than any current pre or post training objective that the field has nailed AlphaEvolve & other LLM + selection loops are also still v short sighted. Science is hard! I think people will have an important role in AI research for at least another decade, & not just due to institutional inertia Recursive self-improvement is obviously going to ramp up before then, but I think it's much more likely that RSI is localized to certain regions of the skill tree Most people that talk about RSI seem to have *foom* AGI in mind? Scoped RSI (e.g. on coding or data analysis) seems much more plausible

@braindersnn Yep, proprietary models will be prolific

My team at NVIDIA partners with lots of startups. I see the opposite happening. RL and post training are really enabling so many companies to take proven ML methods and open models and specialize those models to specific use cases and domains. Always using the biggest model is very slow and expensive. Important when you need it, inefficient when you don’t. Specialization will be an important trend and differentiator in the next wave of AI startups. We are entering a new Cambrian era for lean teams.

@Yuchenj_UW LLMs are not the only form of AI. Even with very strong base models, medicine, robotics, climate, manufacturing, and finance still require domain-specific data, workflows, integrations, validation, and distribution.

Some people at frontier AI labs told me they believe startups are over. OpenAI, Anthropic, Google, xAI will absorb every industry as AGI nears. Coding today, science, medicine, and finance next. Then everything else. If they’re right, that’s a pretty boring end of the world.

@ClementDelangue OS AI is the future. The remaining bottlenecks are compute and good ai r&d

that's why we need open-source AI!

The idea is simple in principle. The engineering is what makes it practical. Paper: arxiv.org/abs/2603.15031 Code: github.com/MoonshotAI/Att…

On a 48B MoE model trained on 1.4T tokens, vs standard residual baseline: - Same compute budget, 1.25× lower loss - Gains across every benchmark: +7.5 GPQA-Diamond, +3.6 Math, +3.1 HumanEval - Less than 2% inference latency overhead

For 8 years, attention only looked sideways. Kimi just rotated it 90 degrees! A thread on Attention Residuals (AttnRes). arxiv.org/pdf/2603.15031