qrdl

@VitalikButerin @shayne_coplan @mansourtarek_ @giancarloMKTS @aosipovich @Polymarket @metaculus @kalshi I'm a huge believer in Prediction Markets. It is free speech, but with accountability. Messengers from the future, warning us of the consequence of our actions. But the markets have significant negative social utility when they are subjective and opaque. It can not and it must not be just about gambling. Prediction Markets have perverse incentives to spread disinfo and manipulate outcomes, so more work needs to be done to surface authentic signal and increase transparency to stop bad faith actors. Polymarket, metaculus and kalshi are the leaders in this industry. Polymarket is arguably the leader and best in terms of transparency and free speech, but has serious problems with bad actors and poorly written subjective markets. Metaculus lack of trading means it doesn't have bad actors, and the rulesets are very well done, but their poor transparency is devastating and destroys a great deal of potential value. They also have poor accuracy and no emergent signal / breaking news because of their minimal incentives. Kalshi has reasonable rulesets, but too many bad actors which is a result of their weak transparency. Through no fault of their own, DCM status suppresses their ability to embrace free speech. I am hopeful, and perhaps even a little optimistic, that 2025 will be the year when Prediction Markets grow up. They will realize their greatest strength is their openness, that they have a profound responsibility to make the world better, and they are not here to just facilitate the transfer of wealth from the naïve and gullible to the sharp and cunning.

@gaoj0017 @grok is the random rotation in turboquant considered a standard technique in optimizing for models?

We need to publicly clarify serious issues in Google’s ICLR 2026 paper TurboQuant. TurboQuant misrepresents RaBitQ in three ways: 1. Avoids acknowledging key methodological similarity (JL transform) 2. Calls our theory “suboptimal” with no evidence 3. Reports results under unfair experimental settings We have expressed our concerns to the authors before their submission, but they chose not to fix them in their paper submission. The paper was accepted at ICLR 2026 and heavily promoted by Google (tens of millions of views). At that scale, uncorrected claims quickly become “consensus.” Facts: 1. RaBitQ already proves asymptotic optimality (FOCS’17 bound) 2. TurboQuant uses the same random rotation step but misses stating the connection 3. Their experiments used single-core CPU for RaBitQ vs A100 GPU for TurboQuant None of these is properly disclosed. We’ve filed a formal complaint and posted on OpenReview (openreview.net/forum?id=tO3AS…). We’ll release a detailed technical report on arXiv. Our goal is simple: keep the academic record accurate. Would appreciate people taking a look and sharing.

The TurboQuant paper (ICLR 2026) contains serious issues in how it describes RaBitQ, including incorrect technical claims and misleading theory/experiment comparisons. We flagged these issues to the authors before submission. They acknowledged them, but chose not to fix them. The paper was later accepted and widely promoted by Google, reaching tens of millions of views. We’re speaking up now because once a misleading narrative spreads, it becomes much harder to correct. We’ve written a public comment on openreview (openreview.net/forum?id=tO3AS…). We would greatly appreciate your attention and help in sharing it.

@gaoj0017 the random rotation is critical and unfortunately they don't alleviate that, however I don't think it was unique to rabitq

@fchollet @fchollet We have a tiny world, it's called math. Open Math problems are a much superior benchmark to arc-agi. Impossible to benchmax (cough), and solving math gets us on the way to solving science.

Many people expect that current AI is ready to cure cancer and do breakthrough new science. ARC-AGI-3 envs are like a microcosm of the scientific method: you must observe a tiny world, form a theory of how it works, test it, iterate until correct. Over the course of a few minutes. If AI can't do it in an ultra-simple, ultra-small scale setting that is explicitly designed to be as accessible as possible, I expect there are a few steps missing until AI can crack the nature of reality.

@FakePsyho What I find stupid in all of this, is we already have extremely useful benchmarks - open math problems. Impossible (mostly) to benchmax. And solving them is the beginning of solving science leading to breakthroughs. Like, wtf, already. Enough with the nonsense.

At this point, I've played through ALL of the games there. I might do a longer write-up describing all of the flaws. Sadly, this would probably get lost in a flood of "AI is 100x worse than humans!!!111" posts from clueless influencers. While it's clear this whole thing is designed in a way that the models will artificially get as low score as possible, I also suspect that game designers / tester team were just not experienced enough to design something fair. I'd expect that the private test games have an order of magnitude more terrible ideas (considering they are expected to be harder).

AI (or any human) will never get 100% in ARC-AGI-3 Let me introduce you to the worst game mechanic you can find in a puzzle game: fog of war At the start, if you go right instead of bottom, you're wasting many moves. Your score on this level literally depends on a conflip!

@lossfunk Can you post an illuminative example? What is sort of the simplest problem it fails on that a human could reasonable do?

🚨 Shocking: Frontier LLMs score 85-95% on standard coding benchmarks. We gave them equivalent problems in languages they couldn't have memorized. They collapsed to 0-11%. Presenting EsoLang-Bench. Accepted to the Logical Reasoning and ICBINB workshops at ICLR 2026 🧵

@mkurman88 i've been using developers.openai.com/codex/cli/ it's pretty scary good

@QRDL I haven't tried it, but I use it in VSC, and it's pretty good there

That's insane, I get more usage quota (or at least can do more) with the $10 GitHub Copilot Pro plan than with the $50 OpenAI Business plan

@DavidColetto Trump is not a villain, he's a symptom.

What stood out most was what he didn’t do. He rejected Trump’s 51st state rhetoric and trade claims. But he did not frame Trump as the central villain. Instead, he argued Canada’s anxiety reflects something deeper: overreliance and domestic weakness. Wrote about the speech: open.substack.com/pub/davidcolet…

@DelongNuna @DavidColetto We gotta stop blaming and start proposing solutions. I respect Pierre for this reset. Hope he keeps it up.

@DavidColetto Pierre NEVER blamed Canadians. He blamed the people responsible for our mess, Trudeau & now Mark Carney. Who is using the same China election propaganda to divide Canada & USA relationship further. Amplifying it for votes.

Pierre Poilievre’s latest speech felt like a tonal reset. For two years, the message has been that Canada is broken. Elites failed. Gatekeepers blocked growth. It was sharp and oppositional. This time, it sounded different. Calmer. More reflective. More like statecraft. Wrote about the speech: open.substack.com/pub/davidcolet…

@4gravitons @martinmbauer @ALupsasca arxiv.org/pdf/2601.22401

@QRDL @martinmbauer @ALupsasca True, but my guess in this case is this isn't a huge player. I don't think there are a lot of OpenAI/scientist collaborations of this type. This one took a personal connection, for example. And the claim seems to be that the twelve hour reasoning run was done in one shot.

Yes, this is a significant result and a solid research paper. And it would’ve been much harder to achieve without GPT. While I understand the instinct, I think it is more interesting to evaluate what type of contribution the AI has made as opposed to focussing on how relevant the result is. ChatGPT generalised a previously derived result for an amplitude that was assumed to vanish for all physical kinematics people care about. These amplitudes are very complicated, lengthy expressions with certain structures and symmetries that are sometimes hidden and difficult to see. This kind of problem is exactly where AI shines! AI is better at detecting breast cancer than a clinician because it has seen millions of scans and detects structures where humans -which are limited by their lifetime exposure- can't. AI that systematically surveys many large amplitudes has a similar advantage Similar to the Erdős problems, it was mostly really an attention bottleneck that left this problem unsolved. This calculation was considered another elaborate way of arriving at zero. So few people were interested in the result and even fewer were working on it. Most if not all physicists would therefore consider the insight of the human physicists that there is in fact a kinematic region where these amplitudes are not zero the most meaningful piece of progress here Both these points aren't to diminish the result. It is seriously impressive and deserves the publicity. It shows where the strengths of modern models can significantly accelerate science and I'm convinced there will be even more relevant discoveries in the future.

@4gravitons @martinmbauer @ALupsasca There was a good paper recently about this

@martinmbauer @4gravitons @ALupsasca But is it survivor / selection bias? We need a detailed analysis of how much time has been wasted analyzing when AI hallucinates and gets it wrong. Stochastically parroting results has real utility, but not if the marginal benefit is eliminated by hallucinations.

@4gravitons I guess it depends how you count, but it should've seen many terms in its training data. It probably didn't see a labelled set (as for mammograms), but on the other hand there're many consistency conditions and symmetries that help. @ALupsasca probably would know better

@nasqret Truth is we are getting very close to that then, at least for me. I'm not sure I can reliably tell which model is the smartest anymore.

@QRDL Singularity

Interesting paradox. With the new Codex-Spark I can generate new content for analysis (documentation, review text etc.) so fast that I cannot parse it and read efficiently. In the end, we humans are the bottleneck in the information pipelines. Thinking slow is essential, but we might be too slow for many tasks. What then?

@lukaszstarosta @AnthropicAI Anyone decent could do this in a day

@AnthropicAI This is terrifying actually. It used to take people months (and their mental sanity) to build something relatively as complex, just 1 year ago. Now we get this with little to no human effort

New Engineering blog: We tasked Opus 4.6 using agent teams to build a C compiler. Then we (mostly) walked away. Two weeks later, it worked on the Linux kernel. Here's what it taught us about the future of autonomous software development. Read more: anthropic.com/engineering/bu…

@0xdoug I am very afraid of a world where people are largely redundant due to automation, but resources are insanely expensive. Very very afraid.

@0xdoug I think the right answer is AI will *eventually* lower prices. But when? And jevon's paradox, will that just shift costs to raw resources, which are becoming increasingly more expensive to extract from the earth. That is my biggest worry in all of this, tbh.

So to recap, AI is expected to create huge amounts of wealth, yet… Workers lose because they’ll be replaced by robots… Software corps lose because they’ll be replaced by Claude code… Tech giants lose because all their free cash flow is now capex… Frontier labs lose because they get margin compressed by open source models that are at most three months behind… Nvidia loses because it gets margin compressed by TPUs and Huawei… So where exactly is the value supposed to accrue?

@0xdoug Just look across the categories of every raw material, like copper, silver, and other industrial metals, and you'll see the same story - declining ore grades, have to dig deeper and harder to get at them. We need massive breakthroughs in material sciences, ASAP.

@0xdoug People often forget there are 8 billion other people on the planet. They have yet to become consumers like the top 13% or so. When that happens (easier when you can just automate everything) - it could put massive strains on raw inputs..

@0xdoug What if we just find out the real problem was never labor and rather it was mining?