In fact qwen knows much better math than their final fixed model. The nonthinking qwen 0.6B parameter model had 55.2% correct on math-500 vs their final improved 48.1% fixed-reasoning model, both with a 32k context window. The original qwen reasoning model had 77.6% accuracy on same test, though qwen recommends using a much larger context for these benchmarks. I wonder if their broken baseline run with 7.9% accuracy is a vllm or setup bug, but didnt dig into it any deeper.
Qwen in reasoning mode is ridiculously anxious and ends up going in circles. It's not just math, it's how it behaves in general.
Others have also fine-tuned Qwen to shake this behavior already, so it's honestly not entirely shocking that Astra could too, but I think this is the first time it was done "by hand" and not via fine-tune.
To be clear, the tuned model here performs worse than the same qwen model performed before RL, which was also not going in circles. It also performs much worse than the RL model that goes in circles. One could argue that the reduction of context may be worth something, but it appears that the final model is worse in all respects than the model at a much earlier stage of training.
You can ask Astra to study 2 or more models and ask it to import their layers to the Qwen model. With this will it not gain more than what it learned with RL?
I am incredibly optimistic on AI, especially for drug discovery. However, the belief that AI drugs will be more likely to pass phase 2 tests than others is either based on the assumption that AI will mostly be used to attack safe indications (me too drugs for targets that already have therapies operating on the same mechanism), or on some belief that perhaps in a distant future the transfer to human biology is better understood by the AI but not documented or shared with humans. Although I think the latter will come true, eventually, people right now tend to forget that because of the extreme cost of human clinical trials, the preclinical candidate has undergone every possible test that regulatory agencies, consultants, the internal company knowledge have suggested. So the failure on a new drug is only related to the “surprise” differences in humans vs all the tests in animals, cells, or in vitro setups that were previously done. Right now AI can help you get to the clinical candidate much faster and with fewer resources than before, but I believe that the translation to success rate improvement has some assumption that is probably not correct in the context of a fixed/novel target. I am open to the possibility that I miss something fundamental and would love feedback.
It's also the case that preclinical candidates developed this way are even harder to translate into human trials, because regulators and drug developers don't know how to build a safe and informative trial given the novelty of the methods that produced the candidate. So you essentially end up redoing a lot of the work that the AI allowed you to skip in order to do the work that remains. Perhaps a future innovation in "explainable" AI could help here, but that such a technology would itself need to be clinically validated in order to solve this problem.
Source: spent 9 years at a AI drug discovery startup. We advanced 8 clinical candidates and every one of them ran into some variant of this issue.
I love the current naming scheme and I am excited to see the astronomy trend expand in the future. I hope to see asteroids, binary stars, and solar systems at some point soon, as well as neutron stars, black holes, and eventually galaxies, and clusters of galaxies.
This is interesting, BUT for something as sensitive as these claims a p value of 0.03 does not cut it. Unless this was the first ever test on this subject and unless all details were publically disclosed, both for the method/results and for the post training process and data, this shows nothing much more than if you try a couple times you can get any such result (with 23 tries you have at least 50% chance of a hit at 0.03). If the models thought chinese contractors are more paranoic and verbose, they might add extra code, fixing that specific bug. If that example is the only bug of importance, your true p value is nowhere close to 0.03 under any sensible analysis.
Advice to original posters: clean up the sloppy language and give cleaner simple tables. But making it to HN, already got you some larger audience than expected, so good job, and good luck with convincing people to use a “fixed” but dated model :-)
As a tourist in SF about two weeks ago I experienced a bit of superhuman driving from a Waymo. On one of these 30-degree downhill slopes, a long truck was stopped in the middle of the road and to its right and slightly ahead was a parked car. The space between these two was just about right so that a 45-degree-turned Waymo car would barely fit with less than an inch space on either side. This is exactly what it did, without much hesitation. As a human driver, I would have stopped, tried to squeeze back up and find another way, or stop and try to locate the truck driver and ask them to move.
> What does robotaxi offer apart from less taxi-driver jobs?
One more passenger seat (once the wheel is removed); the option for privacy (once the recordings are removed); increased safety (starting now, but even more so in the future); ability to build overnight charging stations in unpopulated areas; once fully adapted you can have reduction or removal of parking spaces, eliminate commercial daytime traffic, etc. Maybe it is the ever optimist in me… I think eliminating a job that led to increased cardiovascular problems and reduced lifespan of fulltime drivers while keeping the human rate of accidents, will be seen as an obvious improvement in human civilization in the future, similar to giving up horse carriages.
I put £15.73 (from a dollar exchange conversion of $21.20) at the beginning of September just to try DeepSeek out via API using Opencode and Pi. I still have plenty of credit left! Their off-peak reduced token cost truly is amazing.
These posts came shortly before the math release by OpenAI. I hope the continued discussion of the goldmines in their release helps clarify a vision for how to proceed.
No, why would it be? Unless very detailed web-based microphone or camera manipulation is needed, a modern LLM can built a fallback using safer technology instead.
Not sure why you say that. All the major AI-generated proofs so far were done by AI without using lean, and then later validated by AI formalizing with lean in a separate process. Do you mean we would not have enough rigorous training data perhaps?
Yes that is the case. DeepMind's AlphaProof/Geometry that won IMO Silver used lean as an intrinsic part of the process 2 years ago, but things have shifted. When DeepMind and OpenAI won IMO Gold a year ago, it was solved end to end in language and translated to lean afterwards, and that has been the case since.
because LLMS are like drunk geniuses, without the rigor of lean they would drift into nonsense. Similar in chess they have 1700 ELO now but still some percent of their moves are illegal.
Cool work anyways.
reply