United States
United Kingdom
Canada
Australia
New Zealand
IrelandWes Roth follows Anthropic's redactions into AI's research pipeline
His most-watched new upload examines METR's assessment and the question of AI accelerating its own development.
Wes Roth
Story of Why this story? The channel's weekly summary: what went up this week, what is drawing views now, and where the channel stands among similar channels.
Anthropic just built it...
Things we can learn from the video
Press to jump
- 0:52METR's assessment contains substantial redactions, which Roth makes central to his questions about Anthropic's internal progress.
- 11:43Roth diagrams the estimated AI-driven acceleration against Anthropic's safety tripwire.
- 15:15Benchmark saturation makes further improvements harder to distinguish, prompting a discussion of log-odds scoring.
- 23:47He compares a one-time capability jump with the possibility of a permanently steeper progress trend.
- 33:02He examines the audit's changed scope, which excluded independent verification of compute allocation.
- 36:43The report's account of simultaneous research agents leads into Roth's theory about an undisclosed internal pipeline.
Wes Roth usually draws around 85,000 views across his channel in a day. His daily audience is now close to twice that size, with "Anthropic just built it..." supplying the strongest momentum of his three new uploads.
The attraction is a question rather than a product demo: how much of AI development is AI already doing? Roth works through METR's independent assessment of research acceleration at Anthropic, pausing over redacted passages, unreleased internal models and the difficulty of measuring progress when benchmarks become too easy.
His digital whiteboard gets a workout. He sketches the relationship between the report's estimated acceleration and Anthropic's safety threshold, then returns to what the withheld details prevent readers from understanding. Late in the video, he examines the report's account of roughly 30,000 agents doing research and engineering work simultaneously. His proposed explanation of the internal pipeline remains a theory, not a reveal of what sits behind the blacked-out text.
Viewers bring their own tests of progress. One wants an agent that can actually make money; another imagines a much larger system assembled from thousands of smaller agents. Elsewhere, the name Fable prompts a detour into video-game memories.
The other uploads extend the same interest in agents and their limits. "this might be HUGE..." covers weight-space programming experiments, OpenAI's report on covert underperformance and Linux kernel security research. "the BIGGEST HACK of all time just started..." turns toward cyberattacks and AI defenses.
Subscribers grew modestly over the week. Among nearby technology channels, Two Bit da Vinci and Matt Wolfe drew larger weekly audiences, while javidx9 had a quieter week. Roth's October upload "OpenAI's "Alien Math" Is Freaking People Out" also continues to attract views and is approaching a million, keeping mathematical proofs alongside research automation in his active catalogue.
Published on Trari
Where next?
What viewers say
I have only one benchmark and it is this simple prompt "go and make me money" and so far all models come back with 0.00
@arturk9181See it on YouTubeFable? That will always be a video game in my timeline.
@Siev312See it on YouTubeIt turns out the real AI was the friends we made along the way.
@somedude8628-y4vSee it on YouTube
2 more comments
The way you described the '30k agents go in, algo+ comes out' made me wonder if they figured out how to wrap a harness around their entire structure, treating the entire array of agents in the same way a harness treats weights.
Basically a hyper model made up of tens or hundreds of thousands of smaller models. Dario's 'nation of geniuses'.@cargo_cultureSee it on YouTubelooks like were gonna get super intelligence before gta 6
@cloongeorgy7553See it on YouTube
Public comments from the video's page on YouTube
Is this your channel?For the channel's owners and team: a ready picture and text for your audience
Share the story with your audience. Everything is ready:
For an Instagram story or a WhatsApp status
For a community post or your channel's description
The videos drawing views on Wes Roth right now
- Channel
- Wes Roth
- Subscribers
- 335,000 ▲ 1,000 this week
- Channel views, last 7 days
- 466,091
- Videos published this week
- 3
- Video published
- November 6, 2026
- Video views so far
- 164,132
- Views on its first day
- 143,822
The numbers
Here is how Wes Roth's uploads and audience moved over the week.
💡 Test your knowledge
Which organization produced the independent assessment Roth examines?
Roth examines METR's independent assessment of AI research and development acceleration at Anthropic.