Ah thanks!
The Mistral model has a restrictive license, where you need a specific commercial agreement to use it if you make more than 20M$ revenue/month. Ours is Apache 2.
If you compare the 2, it's not a strict "timestral is better" story - the winrates are tight - your probably should try both on your use-case. We have a 42% winrate vs Timestral on fevbench for instance.
Ofc, our main value is not just these models, but also 1) extra capabilities (like our model that can natively account for stockouts/censored data, and others coming soon), 2) our systems and products around the model, and 3) we're all laser focused on time-series/forecasting
Currently math research serves both as a training pipeline and as a math-production pipeline, and its funding is set based on those two outputs. If it is "just" a training pipeline, do we need as many mathematicians?
Typical thought experiment - go back to the early years of humanity, even before we've created language itself. Communication is basic grunts and other such things, cutting edge knowledge is 'poke him with the pointy end'. Compile all of this knowledge into the training set for an LLM, let it run with infinite processing power and infinite time. It'll just continually remix the knowledge of that snapshot of time. You're certainly not getting to the Moon, discovering the secrets of the atom, or even discovering mathematics.
It sounds increasingly dumb to bring this up given how advanced models have become, but it remains true - they are still fundamentally token prediction algorithms. And that imposes hard constraints on their potential. These constraints are relatively academic in our current world because there's just so absurdly much information and no human can hold more than a negligible amount in his mind at once. So they'll certainly be able to connect many 'obvious' dots that we don't see simply for a lack of carrying capacity, but they're not going to discover concepts that do not clearly follow from the comprehensive knowledge of an era.
It seems far more likely to me that current LLMs have learned generic reasoning capabilities which will allow them to solve effectively infinite problems, rather than whatever vague limit on capabilities you seem to think there are.
I don't get the argument - if we follow what you say, they're not _overreacting_, they are reacting appropriately but should have done so earlier (when o3 was released) rather than now?
My claim is that if Bob, a mathematician, believes that he is rendered useless by an LLM then he did not play with o3 and all the other models between now and then to realize that, yes, they are superhuman at details and deduction but not superhuman at math
AI will not replace the mathematical community (peer review, abstraction creation) until AI can train the next AI purely on data generated by itself and with a tool in a turing complete environment. The environment has to be air-gapped from human data but the AI can use tools and its own outputs
I think that "solving math" is isomorphic to solving the "thinking part" of the scientific process which would be genuine RSI
As other have stated, your point is relevant only for the single best performance point. The question is for interior points of the frontier where the tradeoff actually exists. For instance why is GPT 6 Luna (Max)on the frontier rather than GPT 6 Luna (Medium)? It is not because GPT 6 Luna (Max) is the most performant model regardless of cost.
If training data are purely historical, then how does the human look forward?
Seems to me not impossible that given current knowledge, AI generate one nugget more of knowledge (eg a proof of Navier Stokes), and given current knowledge + the nugget, generate yet some more new knowledge.
I am not sure that is the relevant criterion. The French king as a primus inter pares noble had initially very limited authority over a very limited area. Basically nothing in France was French then, even though they shared State structure and culture, if not strong, direct, central government power.
reply