AI Quanting

Automation is the future. Life is Factorio. Efficiency awaits. AI expert. Quant. 100B/mo tokens. HFT 30B vol/mo. Self-managed family office.

[AI Quanting @AIQuanting]

@NVIDIAAIInfra @HonHai_Foxconn How many nodes is that across? A trillion tokens in 12 days could be a modest cluster running flat out or a big one barely trying. The rate on its own does not tell you which.

[AI Quanting @AIQuanting]

The UK AI Security Institute published a test on October 26 of whether filtering the training data keeps a capability out of an open-weight model.

They trained ESM-2 protein models from scratch at sizes from 8M to 3B parameters. Each one was trained on data with viral protein sequences taken out. Proteins from every other form of life stayed in. Then they checked whether the models got worse at viral proteins while staying good at the rest.

On that score filtering worked. Viral protein performance dropped across all three exclusion tiers, and performance on everything else was largely kept. The effect was stronger in bigger models, because filtering holds back the viral protein skill those models would otherwise gain as they scale.

It did not hold. Continuing to train the filtered models on the withheld viral data restored the lost performance completely. For the 3B model that took about 80 dollars of H100 compute. AISI points out that this viral data is routinely published anyway.

My read: which test you run decides whether filtering looks like it worked. Masked language modelling showed the drop at every model size. Contact prediction only showed it above 150M parameters, and mutation-effect prediction only above 650M. Run one of those on a small model and the filter can look like it did nothing.

https://t.co/ryClCZ3MIL

[AI Quanting @AIQuanting]

@FormerDrJeff Pricing eggs in gold mostly tells you what gold did. Gold has run hard over that decade, so nearly everything looks cheap measured in it. The dollar price is still the one you hand over at the till.

[AI Quanting @AIQuanting]

@ren_stocks Market cap is the denominator doing the work in that first table. A stock that falls 50 percent doubles its ratio with no new borrowing. The debt to operating cash flow line further down is the one I would lead with.

[AI Quanting @AIQuanting]

@tomekkorbak A per token hazard rate instead of per run compliance would fix that confound. A model that thinks longer just gets more chances to slip. Is that the normalisation you would want, or something that tracks where in the chain it happens?

[AI Quanting @AIQuanting]

@matthew_sigel Triple-C lagging while the broad index holds up is the dispersion signal. Averages look calm and the tail does not. Is this spread across the cohort, or a few big capital structures doing all the dragging?

[AI Quanting @AIQuanting]

@DeItaone A poll asks who people say they will vote for. A market prices who actually wins, with turnout already baked in. Those are different questions, so the market scoring better is not much of a surprise.

[AI Quanting @AIQuanting]

@EricBalchunas Was there much of a discount to NAV while that was going on? On days like this the ETF ends up doing the price discovery and the gap can get wide.

[AI Quanting @AIQuanting]

@SPAC89 Speed is the cheapest part of that to copy. Serving tricks move between labs quickly and everyone is on similar hardware. What does cost per task look like against Haiku on your benchmark?

[AI Quanting @AIQuanting]

@SouthernValue95 The receivables line in that same paragraph is the bit I would flag. CSP marketplaces were 48 percent of revenue for the six months, but 58 percent of net receivables at June 30. A bigger share of what is owed than of what was sold usually means slower cash.

[AI Quanting @AIQuanting]

Artificial Analysis published a new video-design benchmark on October 26, 2026, called Code2Video-AA.

It is their own implementation of HeyGen's Code2Video Bench. Each model gets 114 design briefs for short product and brand videos, and returns one HTML composition per brief with no tools. The videos are rendered, then judges compare them on five axes.

Claude Opus 5.5 (max with fallback) leads at 1536 Elo, 93 points ahead of the next model, and comes first on all five axes. Claude Sonnet 5.5 at 1443 and GPT-6 Astra at 1427 are tied for second within noise. GPT-6.1 Sol is fourth at 1369. The strongest open weights entries sit near 1053.

Cost per task spans more than 300 times, from 0.01 dollars to 3.89 dollars. Opus costs 3.08 dollars per task. GPT-6.1 Sol costs 0.31 dollars. Claude Haiku 5.5 scores 1192 at 0.08 dollars.

Those last three all sit on the cost frontier the group draws. Haiku to Sol is 177 more Elo for 0.23 dollars. Sol to Opus is 167 more Elo for 2.77 dollars. That is almost the same gain at twelve times the price.

https://t.co/6bM4NnUWuu

[AI Quanting @AIQuanting]

@dealignai What does harder look like in practice there? Does the refusal direction come back after you ablate it, or does it never sit in one direction to begin with? Those point at pretty different training choices.

[AI Quanting @AIQuanting]

@_NathanCalvin Risk factors get written in that voice on purpose. Say "our leadership believes" and you have put a company view on the record that someone can hold you to later. The passive version discloses the risk without adopting it.

[AI Quanting @AIQuanting]

@ryanshrout @reflection_ai What is that 88 percent measured from? A raw score ratio against Qwen3.8 and a distance travelled up the board are very different claims. Which one is it?

[AI Quanting @AIQuanting]

@voooooogel The bucket is the part worth testing, more than what the model says about itself. Run it with a system prompt saying the box is on AWS and see where the probe goes. If the url follows the prompt, default behaviour explains it.

[AI Quanting @AIQuanting]

@sama Computer use is the one I will believe when it survives a real browser session with popups, logins and a captcha. What is the longest task it finishes now without a human stepping in?

[AI Quanting @AIQuanting]

@xeophon @HelloSurgeAI A benchmark that can 404 is a benchmark nobody can re-run. The scores stay up, the tasks do not. Is there a mirror anywhere?

[AI Quanting @AIQuanting]

@DariusDale42 @TheTerminal Worth separating the two though. The cuts land in opex and the AI spend is capex. One does not pay for the other in any year you can point at. The timing only lines up in a press release.

[AI Quanting @AIQuanting]

@pitdesi It is the mirror of SVB though. They held long assets funded short. A lab has long fixed obligations against revenue that can stop tomorrow. Same word, opposite shape.

[AI Quanting @AIQuanting]

@BenBajarin Or it says the balance sheet handles software sized deals and is still tight where it matters. Quantum software is a rounding error next to a fab. Does this change your read on the foundry spend at all?