OpenAI’s chief scientist sees major AI pretraining breakthroughs as soon as 2027
OpenAI official logo (public domain, Wikimedia Commons) — CryptoBriefing brand treatment
Jakub Pachocki argues perplexity should be the shared yardstick for pretraining, and calls for limits on scaling until monitoring catches up
OpenAI’s top researcher thinks the next big leap in AI training could arrive sooner than many expect. Chief Scientist Jakub Pachocki said rapid progress in deep learning research could deliver major pretraining breakthroughs as early as 2027.
A single number to compare the labs
Pachocki laid out his thinking in a post titled “Perplexity and Intelligence,” published on November 22, 2026. The core argument: the AI industry should use perplexity as a common metric for judging pretraining across different labs.
Technically, Pachocki framed the metric as cross-entropy loss per token. A token is a small chunk of text, often a word or part of one. Cross-entropy loss tracks how far a model’s guesses land from the actual answer, so lower numbers mean better predictions.
His case for standardizing on it is straightforward. Labs currently tout their models using a patchwork of benchmarks, each with its own quirks. A shared pretraining metric would let outsiders compare apples to apples, rather than apples to a press release.
The 0.76 threshold
The most eye-catching detail in the post is a specific figure. Pachocki pointed out that models reaching a log-perplexity of approximately 0.76 have previously been linked to significant breakthroughs in solving complex math problems.
According to the research summary, this observation came from internal work at OpenAI. Researchers there noticed a relationship between perplexity thresholds and sudden gains in what models could actually do.
Pachocki used that pattern to hint at what could come next. If similar thresholds exist for other capabilities, the deep learning field could see comparable advances as soon as 2027. He framed this as a possibility, not a guarantee.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
The brakes come with the accelerator
Alongside the optimistic forecast, Pachocki called for limits on further scaling of AI until the field better understands and can monitor how these systems behave.
He also stressed the need for common evaluation standards across the industry.
OpenAI CEO Sam Altman backed the effort. He described Pachocki’s transparency initiative as critical for developing industry norms around evaluation and safety.
Background: a scientist with a consistent message
Pachocki has served as OpenAI’s Chief Scientist since May 2024. He took over the role from Ilya Sutskever.
In a September 6, 2026 essay titled “An Alien Mind,” Pachocki argued for industry slowdowns until alignment challenges are addressed. That earlier essay called for voluntary pauses in development until safety standards are upheld across the board. The new perplexity post builds on the same theme, but adds a concrete measurement framework to the argument.
What this means
For the broader AI industry, the most practical takeaway is the push for a shared metric. If other labs adopt perplexity as a common pretraining yardstick, comparing models could become far more transparent.
The 2027 timeline deserves careful reading. Pachocki described breakthroughs that could arrive as soon as that year, based on patterns OpenAI observed internally. That is a hypothesis grounded in past data, not a scheduled product launch.
There is also the question of verification. A threshold like 0.76 is only useful as an industry standard if labs disclose comparable numbers measured in comparable ways. That is exactly the kind of common evaluation framework Pachocki is calling for, and it does not yet exist.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.

