@quadrillion_ai @sailresearchco very cool! congrats on the launch!
English
Sasha Krecinic
376 posts
Sasha Krecinic
@clippybasilisk
nothing to see here
Pacific Heights, San Francisco Katılım Temmuz 2017
220 Takip Edilen110 Takipçiler
When you're doing research 20x faster than a single human, every experiment needs its own fast, isolated environment. We're excited to announce our partnership with @sailresearchco to push the boundaries of what's possible with research environments!
Sailboxes have been excellent on both performance and availability, and capable of supporting difficult, real-life research tasks. We're looking forward to making them available as part of Qualia Cloud (currently in private beta — DM for access!)
Sail Research@sailresearchco
Thousands are already in production: hosting agents, running RL rollouts, powering coding + triage workflows. We’re excited to partner with our friends @quadrillion_ai who are already running parallel research experiments on Sailboxes via Qualia Cloud.
English
The first experimental evidence of recursive self-improvement (RSI).
Autoresearching the autoresearch agent for eight days.
The result beats the harness we hand-tuned for two years, on held-out benchmarks: 🧵(1/7)
English
> replicate J-space on GLM 5.2
> train a reward model and run RL to reduce hallucinations
> show me how this model makes cancer predictions
Using our platform Silico is like having a team of AI researchers ready to run experiments like these.
Private beta is open now. 🧵 (1/6)
English
everyone is shipping CAD agents but nobody is evaluating them
so I built CADBench. 308 real designs × 20 categories. real BREP gates on Zoo, Claude, GPT-5, Gemini, Trellis. IRT θ + p5 floor + $/task Pareto.
the results will make you rethink ai and robotics. link in
comments.
adam@adamdotnew
Opus 4.7 appears to be SOTA at agentic CAD design
English
@nikolaj2030 Yeah his public remarks have been confusing so it’s good this clarifies it a bit
English
I'm glad Demis made his thoughts on this topic clearer to the rest of the world. I'm also glad he is advocating for preparing to slow down AI development if necessary.
Demis Hassabis@demishassabis
English
you can now train:
- a 100B-parameter reasoning model
- for 40-turn SWE agent tasks
- in your own coding harness
- for 1000 RL steps
- on just 6 H200 nodes
- in under 2 days
infra co-design is magical
Prime Intellect@PrimeIntellect
verifiers v1 plugs straight into prime-rl for training. We have been using v1 internally for all our production runs. In this run, we train GLM-4.5-Air on ScaleSWE tasks with under-4-minute steps and 35-turn rollouts, completing 1K steps in 2 days on just 6 H200 nodes.
English
@adi_baradwaj What’s the alternative? The error bars would be +\- 50% otherwise, it’s all hypothetical/thought experiment for something with no prior data.
When doing a simulation you have to show what your assumptions through each step are and go from there.
English
Tech diffusion at the frontier and on the other side of the bell curve are different things.
If I’m understanding correctly you’re kinda arguing that a barista not adopting ai is an argument for why frontier tech should be slower.
They’re talking about the critical path for a subset of people who see that it is the critical path.
Two very different rates of progress and adoption.
English
I'm perplexed at why people who I know are smarter than me, including the authors, buy the epistemic strategy used in the AI 2040 scenario, in which, tl;dr, the famous METR time horizon plot predicts eschatological doom.
I don't believe the authors have sufficiently grappled with why the radiologists and programmers still have their jobs, and why baristas haven't been replaced by automatic espresso machines
Or even why we haven't seen more small-scale scheming / reward hacking / misalignment damages due to the use of coding agents across Big Tech, in which we've already granted such agents significant autonomy.
I think if these gaps were more deeply considered they'd be making a more nuanced and less extreme argument.
In the meantime I'm concerned that resources, advocacy and governance are being influenced by their conceptual model.
English
@DKokotajlo Very cool! Thanks for sharing! I think the most interesting benefit is that it aligns interests to “It has to make everything better for it to benefit me.”
English
As Plan A was coming together, I made this diagram to explain to the team why Total Research Transparency seemed so important to me, and why transparency more broadly did. For example, it's very important for preventing concentration of power. (Explanation below)
English
I like that they included many possible plans
Daniel Kokotajlo@DKokotajlo
In AI 2027, we predicted that AI would take over the world or irreversibly concentrate power. In AI 2040: Plan A, we've laid out our positive vision for what should happen instead.
English
@benjamin_horne Click the ai timelines tab
I asked this too
I think the point is it doesn’t matter when takeoff starts, they want to present a variety of timelines. What matters is the mechanics and the forecast and the playbook and the game theory. Don’t anchor too much on exact dates.
English
Not to be a dick but this feels like their sneaky way of hedging their original bet/prediction.
i.e. "If we don't get RSI/fast takeoff/superintelligence in 2027, it's not because we were wrong, it's because things are going according to plan, and we've successfully delayed progress!"
Andrew Curran@AndrewCurran_
The AI 2027 team has written a new scenario, AI 2040.
English
This is an exceptionally well thought-out scenario and a detailed vision for what it might look like for humanity to survive superintelligence.
I endorse aiming for Plan A as a good path forward, but I think we should primarily strive for Plan S.
Daniel Kokotajlo@DKokotajlo
In AI 2027, we predicted that AI would take over the world or irreversibly concentrate power. In AI 2040: Plan A, we've laid out our positive vision for what should happen instead.
English
@jacobrintamaki 💯 totally agree
really makes you appreciate the magic of it + makes sense why they could predict so many things so accurately. Something to be said about detailed work around assumptions and crafting a proper model of the world.
English
in general (not just for AGI essays) there's a lot of stuff that gets thrown into supplements that nobody reads but probably is important for understanding the authors
Daniel Kokotajlo@DKokotajlo
In AI 2027, we predicted that AI would take over the world or irreversibly concentrate power. In AI 2040: Plan A, we've laid out our positive vision for what should happen instead.
English
Eagerly waiting for this to drop. Lots of nuggets in here for policy makers and safety folks.
Very grateful to have gotten a chance to play a very tiny part in it.
Well done @AI_Futures_ team!
Daniel Kokotajlo@DKokotajlo
The AI 2040 scenario: ai-2040.com/?r=tw Plan A is our current best guess, but hopefully a better plan will exist before it's too late. We hope that the best ideas from Plan A will be adopted and the worst ideas discarded. Written with @thlarsen, @romeovdean, @eli_lifland, and @RyanGreenblatt
English
@DKokotajlo @thlarsen @romeovdean @eli_lifland Thanks for doing this for humanity again and for including me in the process ☺️🫡
English
The AI 2040 scenario: ai-2040.com/?r=tw
Plan A is our current best guess, but hopefully a better plan will exist before it's too late. We hope that the best ideas from Plan A will be adopted and the worst ideas discarded.
Written with @thlarsen, @romeovdean, @eli_lifland, and @RyanGreenblatt
English
In AI 2027, we predicted that AI would take over the world or irreversibly concentrate power.
In AI 2040: Plan A, we've laid out our positive vision for what should happen instead.
English
"SAMSUNG IS ALSO PUSHING TO INTEGRATE IT WITH PROCESSING-IN-MEMORY, OR PIM"
The first serious AI PC
Jukan@jukan05
SAMSUNG IS DEVELOPING GAIA, AN AI ACCELERATOR FOR AI PCS. THE COMPANY HAS ALREADY SUPPLIED PROTOTYPES TO MAJOR CUSTOMERS SUCH AS CHINA’S LENOVO AND U.S.-BASED HP FOR PERFORMANCE VALIDATION. MASS PRODUCTION COULD BEGIN AS EARLY AS NEXT YEAR. THE CHIP WILL BE MANUFACTURED ON A 4NM PROCESS. SAMSUNG IS ALSO PUSHING TO INTEGRATE IT WITH PROCESSING-IN-MEMORY, OR PIM, A NEXT-GENERATION DRAM TECHNOLOGY CAPABLE OF PERFORMING COMPUTATIONS ON STORED DATA ITSELF. UNLIKE CONVENTIONAL PROCESSORS THAT SERVE AS THE BRAIN OF A PC, GAIA IS SPECIALIZED FOR AI COMPUTATION. IT IS DESIGNED WITH AN OPTIMIZED NEURAL PROCESSING UNIT, OR NPU, ARCHITECTURE TO EFFICIENTLY HANDLE GENERATIVE AI TASKS ON PCS.
English
@0xBADB01E “Blink once for yes and twice for no…”
“Yes, yes”
“I knew it!”
English
Everyone wondering why I posted about that architecture doesn’t realize that we have something a lot more exciting and consequential in my back pocket.
If I released that it would probably wipe trillions of dollars off of the global stock exchanges. Most likely will never see the light of day as it’s out of my hands now.
English