Aditya Bhattad

Just fine tuning the plasma. Its important to get the vacuum and the current just right for optimal fusing plasma!

@prz_chojecki Damn this is unexpected

Erdos Problem #421: erdosproblems.com/forum/thread/4… Solution I got with GPT-5.6 Sol: ulam.ai/research/erdos…

Seems I was too quick to judge GPT-5.6. In the end I've got to come up with an entirely new construction, not following the footsteps of GPT 5.2, 5.4, 5.5. The Erdos Problem I had used for testing is #421 (link below) and it's about constructing a sequence with distinct products and density 1. The construction that looks promising is a greedy construction, but it turns out it's a nightmare for bookkeeping. GPT-5.6 Sol Ultra came up with a different construction - "gap-greedy" - focusing on primes. This seems to vastly simplify the argument. I post the link to the solution claim below.

@AcerFur Im trying something similar. Im just trying to shave off 246 tho. Building up on some novel results from 5.5 pro and fable.

currently running a long experiment...

@prz_chojecki Shouldnt there be a step for making the right evals (which is basically the right problem selection ig) somewhere in there tho?

RSI should basically be a loop of: - take all the data (e.g. Internet) - enhance with reasoning traces by SOTA model (on-policy rollout) - train - repeat The problem is prohibitive cost of running this loop.

@prz_chojecki Holy hell, a 3b model with imo level reasoning. This is huge and should be recieving a lot more attention.

VibeThinker-3b punches way above its weight class. On our internal IMO-like benchmark it solved 15 problems, while SU-01 30b solved 12. After a short GRPO RLVR with our ulam data it solved 17 on a re-run. Compact intelligence at its best.

@Tesla_Optimus My my

Hand jobs available, come join tesla.com/careers/search…

@prz_chojecki Hoping to see you benchmark sakana on math stuff. Curious to see how it turns out. While it probably wont beat frontier reasoning maybe it can beat frontier at optimization (repeated iteration) and long computations of novel constructions?

This is interesting, because it's the second instance, after OpenRouter Fusion, of an adaptive orchestration API. And I believe not the last, as this category will only grow from here. Directing some tasks to lesser models is something that frontier labs are already doing internally. Doing that cross-company, and basically building your own model for routing seems like an easier way to become a frontier AI lab. And probably not less profitable if you can allow people to complete their tasks while saving on token spend. Potentially this kind of orchestration harness could surpass any single model. That would definitely be true for GPT-5.5 Pro + Fable duo. Both together are better than individual ones, but prohibitively expensive unless you well understand which how to divide and distribute tasks. Very very interesting.

@WeekendInvestng Inflation peddling in the age of ai which is a deflationary tech. If things work out, we should see the price of goods and services fall not rise!

I remember in late nineties when i just started business, it was a dream to achieve 5 cr and then one would need nothing more. 30 years later, 5 cr will not get you a comfortable retirement in any large metro.

@elonmusk Never heard such pessimism about the existence of wormholes before.

In the future, a trillion times a trillion dollars will be spent on making antimatter to travel to other star systems

@alexwg If we extrapolate the complexity theory papers premise to computational biology and materials we get some very interesting predictions. Primarily that there is no general polynomial-cost compiler for arbitary phenotypes.

@rand_longevity 🤞🏼

we are officially on fable 5 watch today

@NoahChrein Bcz it's not an ontology thing right

@datbrownnigga Nobody has the power to yet. At least I don’t

I think its possible to teach set theory, topology, algebra, algebraic topology, differential geometry, category theory, homological algebra and higher category theory in a single hour lecture with zero examples to an audience of 160 iq shape rotators.

@rand_longevity I think they will probably figure out how to fix the jailbreak and re release it and then someone else will obv jailbreak it again. The whole thing reminds me of bane from batman. "Theatricality and deception are powerful agents"

wait was that a bad dream or did we lose fable last night?

@AcerFur @VictorTaelin I dont think gatekeeping intelligence is a good thing. The way to defend against nefarious actors is to employ that same intelligence at scale, not to gatekeep it.

@VictorTaelin Open sourcing such a capable model sounds disastrous. Bad actors would certainly fine tune it to do nefarious cyber-related tasks, and we’d see several open source exploits be utilised in designing sophisticated malware. Irresponsible thing to do; the internet would be in flames.

Futile appeal: Mythos should not be re-launched. It should be open sourced. If it was, that would mark the beginning of an era of unfathomable progress and prosperity to humanity that would make EVERYONE richer, including those at big labs, and the US. I spent the last hour writing a 10 paragraph long text explaining why that's the case and addressing every single point and counter-argument to it, like security, business, economical impact, China - in ways I know to be true, even though hard to see - but I deleted because it became massive and obviously the people that matter would never read it, let alone have a slim chance of being convinced, given that's the extreme opposite of everything they believe. But that is a shame, because it is impossible to overstate how much our species would benefit, if that was true. Things would change in ways that seemingly nobody alive understands, particularly these who should. I suppose our monkey brains are just not smart enough to play the right cards here, and, as a civilization, we all lose. Imagine if an alien species found an infinitely deep oil pit, and used it ALL as a hair moisturizer. We'd laugh at them, wouldn't we? Well, that's us with LLMs, right now. tl;dr way too many people alive today trading a ticket to Mars for a shiny green umber in their phones, and they don't even know that :(

@prz_chojecki Dude I got cut out in the middle of a proof attempt. Now trying to see it through w 5.5 pro 🥲 I got one super interesting analytic NT result with it tho...polishing and reviewing it rn. (Btw pls check your dms!)

Sad to see it go so quickly. Mythos is at least half a notch better than GPT-5.5 Pro and I was already getting some interesting mathematics with it. I'll clean and post what I got with it soon. Maybe this will be a final wake up call for Europe. We need more European models, even if a bit worse than SOTA.

@AcerFur Aaaand just like that tier 4 is saturated now. Dang.

Claude Fable 5 result for FrontierMath T4 has just come in and it is vastly SoTA.

@prz_chojecki @Kimi_Moonshot @GeminiApp @MiniMax_AI @deepseek_ai @NVIDIAAI What are your compute requirements? I'll see if I can help out with this.

TBD, we have the data+architecture ready, we're currently doing various tests with modest compute, before we run it on something bigger (securing enough cheap compute power is also an issue, as we're heavily resource-bounded currently). Would love to see it in 2 months max. It's gonna be OS, only one general model or two - we're targeting 24b/16b parameters. Actually I think it's better not to specialize to particular fields, because you want the model to do cross-field applications.

While everyone talks about Mythos vs GPT-5.5, we've tested other near SOTA models on our ErdosBench. Smoke test on 14 problems with 7 models: Kimi K2.6, Gemini 3.1 Pro, GLM 5.1, MiniMax M3, DeepSeek V4 Pro, Nemotron 3 Ultra and Gemma 3 27b. The winner overall is... Kimi K2.6

@prz_chojecki @Kimi_Moonshot @GeminiApp @MiniMax_AI @deepseek_ai @NVIDIAAI Excited for it! Would love to get my hands on it asap :)

Releasing when? Boy, am I excited for Ulam-1. Rooting for you guys to be releasing the first small LLM capable of doing genuine math research! (I have so many questions. Is it gonna be os? Are you guys gonna release different models for different fields like combinatorics, analytic number theory etc?)

@datbrownnigga @Kimi_Moonshot @GeminiApp @MiniMax_AI @deepseek_ai @NVIDIAAI Definitely on my to-do list, especially that our first model will be a custom fine-tuned SU-01.

@prz_chojecki @Kimi_Moonshot @GeminiApp @MiniMax_AI @deepseek_ai @NVIDIAAI What about SU-01. Might be worthwhile to include in this

8) If you think it would be useful to cover other models in details, let me know! I'll be rerunning it on more problems. Also if you need a private benchmarking for your own reasoning model, DM me! We have benchmarks and can help with crafting custom datasets to boost your model's reasoning.