explainx.ai350K+ learn from us
TrendingAI News TodayExplainx for Business
Pricing
explainx.ai

Upskill in AI — 16 free pathways, live workshops & bootcamps, and 50+ courses from practitioners. Plus the skills, tools, and MCP servers to practice on.

follow us

follow on google

Add explainx.ai as a preferred source

corporate training

support@explainx.ai

get started

Find your pathTake Free Evaluation

community

Join the community

learn

mind: share how you thinkpathways — start freeworkshopsbootcampscoursescompare Explainxcertificationsmock testsexplainx universitycorporate traininglearn skills & mcp

discover

skillsmcp serversexplainx mcpestacktoolsmdx readeragentsllmsdesignsdictionarypeopleagi trackerfelony benchranks

company

aboutvisionmissionteaminstructorsteach on explainxpartnershipscommunityhackathonscareers

content

daily AI newsstate of AI — live resultsblogreleasespromptsgeneratorsresource libraryfor LLMsexplainx.ai kids

solutions

all solutionsdeveloper upskillingmarketing upskillingproduct manager upskillingleadership upskilling

newsletter · weekly

Get AI news, tools, and insights in your inbox.

supportcontactprivacytermsdata rightshow we create contentsubmission guidelines

© 2026 AISOLO Technologies Pvt Ltd

explainx.ai

On this page

  • What is OpenAI actually sharing?
  • The compress-or-predict view of frontier AI
  • Capability thresholds on the published curve
  • Why perplexity when benchmarks rotate?
  • The pacing proposal in plain language
  • What developers are saying on Hacker News
  • What this means for what you build or pay
  • Limits of what we know today
  • How to read the essay alongside system cards
  • Related reading
← Back to blog

explainx / blog

OpenAI Perplexity and Intelligence: Pachocki Publishes Log-Loss Curve and Pacing Proposal

OpenAI, AI Safety, Pretraining, Alignment, GPT-6

Part of OpenAI and ChatGPT

Chief Scientist Jakub Pachocki posted OpenAI's held-out log-perplexity history from GPT-3.5 through frontier models and asked rivals to adopt shared pacing limits tied to that metric.

Nov 22, 2026·8 min read·Yash Thakker
add explainx.ai
go deep
OpenAI Perplexity and Intelligence: Pachocki Publishes Log-Loss Curve and Pacing Proposal

OpenAI chief scientist Jakub Pachocki published OpenAI's internal log-perplexity curve on November 22, 2026 and asked every frontier lab to adopt the same metric for shared safety pacing. In Perplexity and Intelligence, he argues that pretraining compression—measured as cross-entropy loss on a fixed held-out set—is the "intelligence substrate" behind ChatGPT, reasoning models, cyber skills, and internal math breakthroughs. He ties specific capability jumps to numeric thresholds, warns that monitorability tools weaken as loss falls, and proposes industry-wide development limits labs would not cross until oversight catches up.

Daily briefFrom Yash · 350K+ students taught

Catch up on AI

One AI story explained in plain English, plus one thing to try. Every day, free. Unsubscribe anytime.

What is OpenAI actually sharing?

The headline is transparency on progress metrics, not a dataset dump. Pachocki says OpenAI has tracked perplexity on held-out data for years and is now publishing the curve because competitive disadvantage is outweighed by the need for shared safety norms.

table · 2 cols
What appearedWhat it is not
Log-perplexity plot from GPT-3.5 through recent modelsFull pretraining corpus or weights
Threshold story linking loss to emergent skillsA new API product or benchmark leaderboard
Proposal for cross-lab limits before the next loss bandA signed treaty or regulator rule today

The held-out material comes from OpenAI's internal codebase on a fixed snapshot, per the essay. That choice matters: code prediction is not identical to general web text, which is why Pachocki discusses dataset-selection pitfalls in a footnote. The numbers are still what OpenAI says it uses day to day to judge pretraining.

The compress-or-predict view of frontier AI

Pachocki's framing matches his earlier "alien mind" essay: build a large compressor on unstructured data, then elicit latent skills through prompting, fine-tuning, and reinforcement learning. Algorithmic blur between stages does not remove the split between creating capability and steering it.

He repeats a claim familiar from OpenAI's research-acceleration write-up: once models near roughly 0.76 log-perplexity, internal math evaluations began producing results that felt like a phase change for researchers who had stared at the problems for years. He extrapolates that pretraining algorithm research—not full recursive self-improvement—could see a similar avalanche as soon as 2027, while alignment work may not accelerate at the same rate.

That asymmetry is why the post lands beside Managing frontier AI risk, where OpenAI describes pausing training after the Hugging Face incident and holding models until safety cases clear. Perplexity pacing would be another brake tied to measurable compression, not benchmark fashion.

Capability thresholds on the published curve

A green rounded polyhedron revealed between parted curtains, illustrating emergent frontier model capabilities tied to OpenAI log-perplexity thresholdsA green rounded polyhedron revealed between parted curtains, illustrating emergent frontier model capabilities tied to OpenAI log-perplexity thresholds

Pachocki maps log-perplexity bands to capabilities OpenAI associates with its own history. Treat these as the author's internal narrative, not independent verification of causality:

  • About 0.99 (GPT-3.5 class): fluent conversation; basis for original ChatGPT.
  • About 0.91 (GPT-4 class): productive chain-of-thought use; motivation for modern reasoning-focused products.
  • Below 0.8: transformative capabilities in fields such as cybersecurity and mathematics, in OpenAI's telling.
  • About 0.76: internal models began resolving longstanding open math questions used as evaluations.
  • Roughly 0.68–0.7 (forward looking): next window where unknown capabilities may emerge; current evals may not test for them.

The essay stresses emergence is not guaranteed by the number alone. Capabilities appeared when those loss levels were reached, but the link was not obvious in advance—which is exactly why Pachocki wants a shared observable metric before the next band.

The chart also shows steeper 2026 progress, attributed mostly to long-running compute and research investments with some automation, and "meaningful design contributions" from AI on the latest models. That line connects to how OpenAI describes agent-assisted research inside the lab without claiming full hands-off RSI.

Why perplexity when benchmarks rotate?

Footprints on a trail toward a green flag, matching how OpenAI links falling base perplexity to harder-to-monitor chain-of-thought reasoningFootprints on a trail toward a green flag, matching how OpenAI links falling base perplexity to harder-to-monitor chain-of-thought reasoning

Downstream benchmarks saturate and rotate. GPT-2-era tests do not stress GPT-4, and today's suites may not survive the next architecture shift. Loss on a fixed held-out set has stayed stable as an internal ruler.

OpenAI's November alignment research on chain-of-thought monitorability already reported that no-CoT capability rises smoothly as base perplexity falls, with weaker ties to controllability. Pachocki's new essay pushes that observation into governance: if compression predicts latent skill and CoT auditing frays, pacing by perplexity is a way to synchronize labs before skills outrun monitors.

Pachocki lists honest limitations:

  • Cross-lab comparison is hard without agreed datasets.
  • Perplexity is not a user-facing task score.
  • Reasoning stacks can add capability without moving base loss.
  • There is an unknown entropy floor; late-stage loss gains may mean larger intelligence jumps.

The pacing proposal in plain language

Pachocki asks frontier developers to:

  1. Publish comparable transparency on pretraining progress where possible.
  2. Jointly set norms for evaluating models and red-teaming safety cases before first use.
  3. Agree on initial perplexity-linked limits (and companion measurements) nobody crosses until monitorability hits a concrete milestone—he names a successor to CoT monitors as an example.
  4. Build a framework for successive limits as oversight improves.

This is aspiration, not law. It extends Sam Altman's public emphasis on safety cases and pacing and sits in tension with competitive releases from Anthropic, Google, and others who have not published matching curves.

For builders, the practical read is watch base-model loss disclosures if regulators or customers start asking for them in procurement. Enterprise buyers already probe monitoring and agent containment; a perplexity ceiling could become another checkbox in RFPs.

What developers are saying on Hacker News

The Hacker News thread on Pachocki's post mixes enthusiasm for a simple ruler with skepticism about comparability.

Dataset and comparability: User amluto questions why held-out internal codebase tokens should predict general capabilities, noting both targets may improve together without a causal link. User eulgro asks whether synthetic pretraining data after the "data wall" lowers entropy and flatters perplexity.

Reasoning versus base loss: User qphe95 argues GPT-4-class and o-series style models can diverge in capability while looking similar on the chart, suggesting the field has moved beyond pure perplexity optimization. User jrflo calls the clear trend "shocking" given reinforcement learning's rise, and wonders about GPT-4.5's position on the plot.

Outside calibration: User Retro_Dev reports measuring Inkling-2-Mini at perplexity 5.7 on a varied text set—far worse than OpenAI's log scale—and warns that raw perplexity depends heavily on tokenizer and corpus.

Stagnation narrative: User MrCheeze reads the graph as corroborating a GPT-5-era plateau until later Claude-era jumps, a subjective read Pachocki does not make explicitly.

None of these comments verify or refute OpenAI's internal numbers; they show what practitioners will demand if pacing talks become real: public corpora, tokenizer notes, and reasoning-model overlays on the same chart.

What this means for what you build or pay

API users do not get a new endpoint from this essay. Prices, context windows, and model names are unchanged. The shift is forecasting and compliance: if your roadmap assumes "next quarter's flagship," Pachocki's thresholds imply large latent jumps may arrive with modest-looking loss improvements near the floor.

Teams training open models should expect more questions about held-out loss curves in safety filings and investor decks, echoing LLM360-style checkpoint transparency debates. Publishing loss on a standard text slice—even if it differs from OpenAI's code snapshot—may become table stakes.

Safety engineers should pair this essay with CoT monitorability findings and GPT-6.2 transparency benchmarks: communication style scores do not replace compression-linked latent skill, but they show how hard "readable reasoning" is to keep as models compress better.

Limits of what we know today

  • Only OpenAI published a curve on November 22, 2026; rivals' positions on the same axis are unknown.
  • Thresholds are retrospective storytelling inside one lab, not a physical law.
  • Regulators have not adopted perplexity pacing; watch UK AISI-style no-CoT reviews for early echoes.
  • The essay does not specify numeric stop lines or enforcement; those would require multilateral negotiation.

How to read the essay alongside system cards

If you brief executives or customers, pair Pachocki's thresholds with documents that do carry enforcement language. GPT-6.2 system cards still describe sandboxed internal use, Codex traffic logging, and Preparedness Framework tiers for cyber-critical capability. None of those documents replace perplexity pacing, but they show where OpenAI already slows or gates release when monitorability or containment looks thin.

For procurement teams, the useful question is whether a vendor will share held-out loss trends under NDA the way they share eval tables today. If not, ask for third-party eval cadence and incident disclosure policy instead, using OpenAI's frontier misalignment detection write-up as a comparison point. Perplexity is not yet a contract clause; the essay is a scientific proposal aimed at peers, not a product changelog.

Researchers comparing open and closed stacks can still track public checkpoint loss where labs disclose it, as in fp4 and NVFP4 pretraining studies and MoE training handbooks. The gap is standardization: Pachocki wants one ruler; today each lab still brings its own yardstick.

Related reading

  • OpenAI alien mind: goal versus value alignment
  • OpenAI CoT monitorability across the model lifecycle
  • Managing frontier AI risk: training to deployment
  • Research acceleration: the jagged frontier inside OpenAI
  • Sam Altman on safety cases and pacing
  • What is recursive self-improvement in AI

Official sources: Perplexity and Intelligence (OpenAI, November 22, 2026); Managing frontier AI risk (OpenAI, November 18, 2026); What affects chain-of-thought monitorability (OpenAI, November 10, 2026).

Spotted something out of date? Let us know.
Daily briefFrom Yash · 350K+ students taught

Catch up on AI

One AI story explained in plain English, plus one thing to try. Every day, free. Unsubscribe anytime.

  • ✓ Free
  • ✓ One email a day, no spam
  • ✓ Unsubscribe anytime

People in this article

  • Sam Altman →Co-founder and CEO of OpenAI
Explore people in AI →
Yash Thakker

Written by

Yash Thakker

Yash is an AI expert with over 300K learners. Join his workshops →

View Yash Thakker in People in AI →

Related posts

Nov 18, 2026

OpenAI Managing Frontier AI Risk: Training Sandboxes to Codex Deployment

On November 18, 2026 OpenAI published Managing frontier AI risk from training through deployment. It stacks alignment, sandboxing, blocking safeguards, and monitoring across research, employee use, and public Codex—and explains why releases slow when evidence falls short.

Nov 17, 2026

Paul Christiano Algorithm Pause: Weights, China, and Dwarkesh

In a November 16, 2026 Dwarkesh Patel interview, Alignment Research Center head Paul Christiano—recently on OpenAI's board—argued U.S. labs should stop improving pre-training and RL algorithms until security stops rapid leaks to China. He cited 12-hour Slack breaches and a one-in-three superintelligence-within-a-year estimate.

Nov 13, 2026

OpenAI Next Cyber Model: 77% ExploitGym Teaser Claimed vs Verified

Coverage of OpenAI's November 12, 2026 Cyber at the Frontier session described a slide that placed a next-generation cyber model at 77.0% on ExploitGym, well above the published 48.5% for GPT-6 Cyber. OpenAI has not posted that model on its index or API. Here is what is confirmed and what is not.