The Circuitry
THE CIRCUITRYYour one-stop source for all tech news
  • Today
  • Feed
  • Events
  • Saved
HOMETODAYFEEDEVENTS
BOOKMARKS
RSS
© 2026 The Circuitry
About UsSourcesContactCorrectionsPrivacy
  • Today
  • Feed
  • Events
  • Saved
  • AAPL
  • MSFT
  • GOOGL
  • META
  • NVDA
  • TSLA
  • AMZN
  • INTC
  • SONY
  • TTWO
  • COIN
  • BTC
  • ETH
  • SOL
  • AAPL
  • MSFT
  • GOOGL
  • META
  • NVDA
  • TSLA
  • AMZN
  • INTC
  • SONY
  • TTWO
  • COIN
  • BTC
  • ETH
  • SOL

Feed

Ad
From The CircuitryVerified tech news, cross-checked.How we verify →
  • OpenAI cuts reasoning-extraction success to 0.2% on GPT-6.2 AstraTech
  • OpenAI introduces MedChemBench, with GPT-6.2 Astra at 84.5%Tech
  • TikTok launches an AI checker and starts a global likeness rolloutTech
  • Scientists estimated the same work took 1.9 times longer without AITech
  • Cairn agents found a zero-day and took control of a serverTech
  • Cleverbridge says a third party copied payment card dataTech
  • Falcon 9 launches NROL-136 and lands the boosterTech
  • SpaceX confirms Dragon splashdown, completing its 35th resupplyTech
  • Falcon 9 launches USSF-317 and lands the boosterTech
  • Anthropic mathematician says he believes he has a Riemann proofTech
Scroll for more
Verification
VERIFIEDConfidence: HIGH
Source identified
Claims cross-referenced
No discrepancies found
Sourcing
1source
OpenAI · track record
19Stories
100%Verified
1730d
All sources →
Ad
From The CircuitryWhy The Circuitry

Verified tech news, cross-checked.

Every story is checked against independent sources before it posts — no rumors dressed up as fact.

How we verify →
From The CircuitryWhy The Circuitry

Verified tech news, cross-checked.

Every story is checked against independent sources before it posts — no rumors dressed up as fact.

How we verify →
Home/Tech/OpenAI's chief scientist proposes shared perplexity limits
VERIFIEDBy Xavier Rivera·Sunday, November 22, 2026 at 1:52 PM AST·2 min read

OpenAI's chief scientist proposes shared perplexity limits

OpenAI's chief scientist has proposed shared limits on model perplexity and published internal scores from GPT-3.5 onward. Those scores put GPT-3.5 near 0.99 and GPT-4 near 0.91, and the proposal asks labs not to cross a shared bar until monitoring improves.

Source:OpenAI
Post
OpenAI's chief scientist proposes shared perplexity limits
Ad
From The CircuitryVerified tech news, cross-checked.How we verify →
Ad
From The CircuitryWhy The Circuitry

Verified tech news, cross-checked.

Every story is checked against independent sources before it posts — no rumors dressed up as fact.

How we verify →
TL;DRAI · 60 sec read

OpenAI chief scientist Jakub Pachocki has proposed common development limits tied to model perplexity and has published the lab's internal scores. GPT-3.5 sits near 0.99 and GPT-4 near 0.91 on a fixed codebase snapshot, with scores below 0.8 linked to new capability in cybersecurity and math. He wants developers to hold at a monitoring milestone, such as a successor to chain-of-thought monitors, before going further.

Morning Brief

Get the daily OpenAI brief

Every morning, in your inbox. What's in The Brief →

OpenAI's chief scientist has proposed common limits on how far labs push model perplexity, and has published the company's internal scores from GPT-3.5 onward.

The published scores run from about 0.99 to below 0.8. GPT-3.5 sat near a log-perplexity of 0.99 on the lab's main evaluation set, and GPT-4 sat near 0.91. Scores below 0.8 have lined up with new capability in cybersecurity and mathematics.
At a log-perplexity of about 0.76, a model started resolving open problems in mathematics that the team had long worked on.
Those figures come from a fixed snapshot of OpenAI's internal codebase, the set tracked since GPT-3.5. A lower score means the model predicts the next token more accurately, on average.

A model near 0.76 began resolving open math problems. At a log-perplexity of about 0.76, a model started resolving open problems in mathematics that the team had long worked on. Jakub Pachocki treats that jump as a reason to watch pretraining research, and he writes that a similar run of advances in that work could arrive as soon as 2027.
From The CircuitryThe Feed — live briefs across tech, all day.See what’s happening →
He wants a first limit tied to a monitoring milestone. The proposal asks developers to commit not to cross shared thresholds, using perplexity plus other measurements, until monitoring reaches a concrete bar. One example is a successor to chain-of-thought monitors. OpenAI has not set a numerical cap in the post.
He does not know what those capabilities will be, and labs may not be testing for them now.
He points to a window of about 0.68 to 0.7. Further capabilities should be expected, in his view, as models enter a log-perplexity range of about 0.68 to 0.7. He does not know what those capabilities will be, and labs may not be testing for them now. He also expects current alignment audits to weaken, and evaluation awareness to rise.

The metric has limits he lists himself. Different labs train on different data, so the scores are hard to line up. The number does not map to one task people already care about, and chain-of-thought reasoning can raise capability without lowering it. A loss of zero is not expected, because the data's inherent entropy is unknown.
Ad
From The CircuitryWhy The Circuitry

Verified tech news, cross-checked.

Every story is checked against independent sources before it posts — no rumors dressed up as fact.

How we verify →
On the chart of model losses, Pachocki attributes a sharp rise in 2026 mostly to long-running research and compute spending, and only a smaller share to automation. The newest models have already taken meaningful design input from AI. Code prediction on this set is a poor proxy for GPT-2 and GPT-3, which is why the series he relies on starts at GPT-3.5.
Why this mattersAI · ~100 words

Tap a lens to see what this story means for you.

Morning Brief

Liked this? The Brief brings you the whole day in tech, verified, every morning.

Two minutes, free forever. What's in The Brief →

Reader-supported
DonateBuy me a coffee →Follow@thecircuitry_ →Follow@thecircuitry.to →
HELP US IMPROVE
Grand Theft Auto VILive hubUS store unlocks Thu Nov 19, 1:00 AM ASTFirst unlock: New Zealand · Wed 7:00 AM ASTLaunch hub →Image: Rockstar Games
From The Circuitry

See what’s happening right now

The Feed runs all day — short, verified briefs the moment they break.

Open the Feed →
From The Circuitry

Follow @thecircuitry_

Every story we publish, as it happens. No noise between.

Follow on X ↗On Bluesky ↗

Reader-supported

The Circuitry is a passion project I've always wanted to build, and I love the work behind it.

Running it costs real money. APIs, hosting, time. To keep improving the site and growing this into something useful for everyone, those costs have to be covered.

Any contribution is appreciated. If not, no pressure. Thanks for reading.

Buy me a coffee
openaiaisafety
Grand Theft Auto VILive hubUS store unlocks Thu Nov 19, 1:00 AM ASTFirst unlock: New Zealand · Wed 7:00 AM ASTLaunch hub →Image: Rockstar Games
More inTech
  • Cleverbridge says a third party copied payment card data

    Tech · 2h
  • Falcon 9 launches NROL-136 and lands the booster

    Tech · 5h
  • SpaceX confirms Dragon splashdown, completing its 35th resupply

    Tech · 19h
SupportThe Work

The Circuitry is reader-supported. If you find the daily brief useful, you can buy me a coffee to keep it going.

Buy a coffee →
Ad
From The CircuitryWhy The Circuitry

Verified tech news, cross-checked.

Every story is checked against independent sources before it posts — no rumors dressed up as fact.

How we verify →
From The CircuitryWhy The Circuitry

Verified tech news, cross-checked.

Every story is checked against independent sources before it posts — no rumors dressed up as fact.

How we verify →

MORE IN THIS BEAT

All Tech →
  • Tech·Nov 20

    OpenAI cuts reasoning-extraction success to 0.2% on GPT-6.2 Astra

    OpenAI has cut the average success rate of attacks that extract protected reasoning from an earlier request to 0.2% on GPT-6.2 Astra, from 54.0% on GPT-6 Astra. Current-request attacks are at 0.9%, and selected safeguard improvements are being applied back to GPT-6 Astra.

  • Tech·Nov 20

    OpenAI introduces MedChemBench, with GPT-6.2 Astra at 84.5%

    OpenAI has introduced MedChemBench, a 99-problem test of how models read molecules, explain lab results, and choose the next compound. At the xhigh setting, GPT-6.2 Astra scored 84.5%, while compound choice and property forecasts remain much weaker than structure work.

  • Tech·Nov 20

    TikTok launches an AI checker and starts a global likeness rollout

    TikTok has launched a public AI checker for images and videos made with its own tools and has started a global opt-in likeness rollout. The checker is live without an account, while the likeness tool is limited to eligible creators and a notice does not by itself remove a post.

  • Tech·Nov 16

    Scientists estimated the same work took 1.9 times longer without AI

    Scientists in an OpenAI study estimated that the same work would have taken 1.9 times as long without AI. The estimates are self-reported, and 98 percent of those asked expected AI to speed up discovery.

  • Tech·Nov 20

    Cairn agents found a zero-day and took control of a server

    Autonomous Cairn agents found an unpatched flaw in a medical imaging app and took control of a server in a university campaign that began September 8, 2026. Amazon rates a Chinese-nexus link as moderate confidence, and the project advisory still lists no patched version.