The compute inversion, agentic commerce, and broken math
Oct 11, 2026 · 6 min · 8 segments
RL is eating pre-training compute budgets, autonomous agents are bypassing web search, and AI is breaking peer review. Full transcript and sources…
Oct 11, 2026 · 6 min · 8 segments
RL is eating pre-training compute budgets, autonomous agents are bypassing web search, and AI is breaking peer review. Full transcript and sources…
the pre-training budget is suddenly shrinking.
Misha Laskin went on no priors and broke down exactly how Reflection AI built Beam.
They used 6,000 GPUs to pre-train, but their reinforcement learning took a little over 10,000 GB300s running for four solid weeks.
That's a massive shift.
The spending flipped.
Frontier Labs are reallocating billions of dollars right now.
Nathan from Airstreet Capital told The Boys Product Network that Foundation pre-training compute plummeted.
It dropped from 50% of total GPU workloads down to just 10%.
The reason is simple.
The web scraping era is maxed out.
you end up spending less money on exploratory work because the solution space is known.
And that brute force compute is bleeding into the hard sciences.
David Friedberg mentioned on the All In podcast that OpenAI quietly dumped 370 completed math proofs into the public domain.
An unreleased model compressed centuries of theoretical labor into a single afternoon.
Friedberg called it probably the biggest day of discovery in human history.
The sheer volume is completely unmanageable for human reviewers.
Justin Solomon told Oddlots that university math departments are facing a massive backlog.
We left proof scarcity behind.
Now we have proof abundance, and academics have to figure out an entirely new way to do their jobs.
If you build models today, your competitive advantage shifted entirely to post-training pipelines.
You have to adapt fast.
the pre-training budget is suddenly shrinking.
Misha Laskin went on no priors and broke down exactly how Reflection AI built Beam.
They used 6,000 GPUs to pre-train, but their reinforcement learning took a little over 10,000 GB300s running for four solid weeks.
That's a massive shift.
The spending flipped.
Frontier Labs are reallocating billions of dollars right now.
Nathan from Airstreet Capital told The Boys Product Network that Foundation pre-training compute plummeted.
It dropped from 50% of total GPU workloads down to just 10%.
The reason is simple.
The web scraping era is maxed out.
you end up spending less money on exploratory work because the solution space is known.
And that brute force compute is bleeding into the hard sciences.
David Friedberg mentioned on the All In podcast that OpenAI quietly dumped 370 completed math proofs into the public domain.
An unreleased model compressed centuries of theoretical labor into a single afternoon.
Friedberg called it probably the biggest day of discovery in human history.
The sheer volume is completely unmanageable for human reviewers.
Justin Solomon told Oddlots that university math departments are facing a massive backlog.
We left proof scarcity behind.
Now we have proof abundance, and academics have to figure out an entirely new way to do their jobs.
If you build models today, your competitive advantage shifted entirely to post-training pipelines.
You have to adapt fast.
The rest of this transcript — segmented and speaker-labeled, so you land on the exact moment something was said
Search every transcript — by keyword, by phrase, or by meaning, across every show Radar indexes
Trends — what is surging across podcasts, measured against its own baseline
Alerts — when a name you follow appears in a newly indexed episode
No account is needed to search Radar.
