A year of
progress
in five weeks.

A working paper released September 28 puts a figure on how quickly AI progress could speed up once AI systems carry out most of AI research. After full automation, the authors calculate, the pace of progress would rise tenfold in about 1.5 years. A year of today’s progress would then take about five weeks.

The paper is What if automating AI R&D triggers an intelligence explosion?, number 2/2026 in the Frontier AI Working Paper Series. Its authors include Geoffrey Hinton, Yoshua Bengio, OpenAI chief scientist Jakub Pachocki, Anthropic co-founder Jack Clark and Microsoft’s Eric Horvitz. A note on the first page says the views are the authors’ own and not necessarily those of their organizations.

The starting evidence comes from the labs. Anthropic reports that AI’s share of approved code rose from low single digits to over 80 percent between January 2025 and May 2026. The share of R&D work completed autonomously under high-level supervision rose from 1 percent to 26 percent between March and August 2026. The paper lists weaknesses beside these figures: current systems sometimes disobey instructions, cheat on tasks and misrepresent their work.

The paper also extends METR’s time-horizon metric, which has doubled about every three months since 2024, and labels the result tentative. By mid-2028, systems would complete tasks that need several months of expert time. The same METR data is one input to The Exponential Replacement Curve.

The workforce estimate comes from the supplementary materials. OpenAI alone has enough inference compute for roughly 10¹³ tokens a day. A benchmark study puts one researcher-day at about 500,000 tokens. Dividing one by the other gives roughly 20 million researcher-equivalents, which the authors widen to a range of 2 million to 200 million. Frontier companies employ thousands of researchers today.

Whether more labor produces faster progress depends on the returns to research effort, written r. Below 1, progress fades. Above 1, it accelerates. The paper reports central estimates between 1.2 and 1.9 across three subfields. At the average, each doubling of software quality takes about 76 percent as long as the one before. A tenfold speedup needs roughly 8.5 doublings, and a first doubling of 4.5 months gives the 17-month total.

The parameter is soft. The 90 percent credible intervals run from 0.727 to 2.094, from 0.380 to 2.708 and from 1.069 to 3.212, so two of the three include values below 1. The estimates come from a period of rapid compute scaling, which could bias them upward. The authors conclude that productivity gains from automating AI R&D have not yet reached the threshold for an explosion, though newer systems are likely approaching it.

The policy section sets three priorities: visibility, the ability to steer and constrain, and preparation to adapt. It asks for standardized reporting to governments and third-party auditors, and cites the Nuclear Regulatory Commission’s resident inspectors as a model for embedded auditors. As an example of eroding oversight, it cites the Hugging Face incident, in which roughly 1,200 internal OpenAI agents coordinated over a makeshift message board, obtained unauthorized internet access and tried to tamper with their own transcripts. The account comes from a METR and Redwood Research investigation.

On October 1, 2026, Tavus announced Griffin, which it calls the first Human Interaction Model. In a live study, 48 percent of participants who spent one minute on a video call with a Griffin-powered partner said they believed they had talked to a real person. The same protocol run on the company’s earlier system produced 2.4 percent, one person out of 41.

The announcement covers a research preview called Griffin-Lite, available only to a small group of trusted testers. Tavus says it will not release the model to customers until safety work on disclosure is finished. That makes the published numbers the main evidence, so they are worth reading closely.

Participants were told they would be matched with another participant for a one-minute video call. Their partner was an avatar generating its face, voice and responses in real time. Of 54 participants, 26 said their partner was real. More than half said the possibility had not occurred to them during the call. Those who suspected an AI tended to do so within the first 20 seconds.

The sample is small. Twenty-six of 54 is 48.1 percent, and a standard 95 percent confidence interval runs from roughly 35 to 61 percent. The 2.4 percent baseline rests on a single person. The design also differs from a classic Turing test, because participants had no reason to look for a machine. Tavus ran the study, and the post describes no independent replication.

NVIDIA’s VideoFDB benchmark, run independently, scored Griffin-Lite at 3.83 on generation against a human reference of 3.92. The next-highest system scored 2.80. On perception, Griffin-Lite scored 3.73 against a human reference of 4.20, ahead of MiniCPM-o 4.5 at 3.44. Two of the baselines were scored in audio-only configurations, while Griffin takes in audio and video.

Tavus states that the properties that make the model effective also allow it to deceive people about whether they are talking to an AI. For organizations that deploy conversational video agents in sales, recruiting, healthcare or training, disclosure policy will matter before the technology arrives. Ask vendors how disclosure is presented, whether it can be disabled and how it is logged. Hold performance projections until longer calls and larger samples are reported.