Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
Deepseek is a Problem
TL;DR
Berman says US open-source AI is 'almost certainly doomed' under current incentives — in America, labs pay the R&D and GPU bill, then rivals can serve the same model with better margins, which breaks the business model for startups.
China’s edge isn’t talent, it’s structure and subsidy — he argues the CCP can back companies to give strong models away cheap, kill margins for leaders, and win enterprise adoption with 'good enough' AI at a fraction of the cost.
Most businesses don’t need frontier intelligence, they need cheap, controllable models — for 'spreadsheets, coding, and making a schedule,' a DeepSeek-class open model is compelling versus expensive proprietary systems from OpenAI or Anthropic.
Nvidia may be the only US company with a real open-source AI business model — Berman points to Nvidia’s reported $26 billion open-source push, arguing it can give models away because everyone serving them still buys Nvidia chips.
Building the US economy on Chinese open models creates strategic risk even if we self-host them — his concern is that China could shape AI standards, chip optimization, and even subtle cultural defaults, while also undercutting revenue for US frontier labs.
The counterargument is Anthropic-style 'straight shot to AGI,' but Berman says that’s too uncertain to bet everything on — if AGI takes longer than hoped, widespread dependence on Chinese open source could weaken the very US compute-and-revenue flywheel needed to win.
The Breakdown
The opening thesis: no middle ground in AI
Berman comes in hot: the US is either 'screwed' or 'going to win everything,' and he frames AI as a national bet tied to the stock market itself, with 40% concentrated in seven tech companies. His core alarm is that open-source AI matters enormously, and the US currently has no durable way to make it work.
Why open source is valuable — and why it’s financially broken
He quickly defines open source in practical terms: the recipe and usually the weights are released, so anyone can download, fine-tune, and run the model. That openness makes models more secure, more efficient, and easier to improve — but it also creates a brutal economic trap where the original builder spends months and millions training a model only to watch others sell it more profitably.
Why China can play this game differently
Berman says China’s advantage is not superior tech or talent, but government structure: the CCP can effectively pick winners and subsidize strategic industries. His argument is simple and sharp — if you’re behind, giving away a very good product for free is a great way to crush the margins of whoever is ahead.
The enterprise decision point favors DeepSeek-style models
He stresses that this moment matters because US companies are choosing their AI stack right now, and enterprise adoption moves slowly. Most businesses are not solving frontier science problems, he says; they need AI for coding, spreadsheets, and scheduling, so if DeepSeek is almost as good and dramatically cheaper, the choice gets pretty obvious.
The weak US open-source bench — and Nvidia as the white knight
Berman runs through the field: Meta talked up Llama and then backed away, OpenAI’s GPT-OSS feels like a side quest, Anthropic has 'zero' open-source strategy, and Google’s Gemma is good but not built to run companies. Nvidia is the exception: with its own revenue engine, top researchers, and incentive alignment, it can lose money on models because everyone who serves them still buys Nvidia hardware.
Why Chinese open models are a geopolitical problem even when self-hosted
He anticipates the obvious objection — if the models are open, why not just run them ourselves? His answer is that the real issue is control over standards and infrastructure: if US enterprise standardizes on Chinese models, China can shape chip design, optimization paths, and broader industry direction, especially since export controls already pushed it to find algorithmic efficiencies on domestic hardware.
The Anthropic counterargument: none of this matters if AGI arrives first
Berman gives real airtime to the opposing view, especially Dario Amodei and Anthropic’s 'straight shot to AGI' logic. If one lab hits recursive self-improvement first, he says, that winner could make cheaper models, discover efficiencies, and effectively own the board — making today’s open-source battle look temporary.
His fixes: treat US open source like infrastructure
Still, he doesn’t buy AGI timing as a reason to ignore the problem, because dependence in the interim could wreck the US flywheel. His proposed menu is practical: federal grants or compute quotas for open-source labs, procurement guarantees in sectors like defense and healthcare, more hardware-funded model development from AMD and Intel, vertical open models for legal/biotech/code/defense, and shared standards to reduce startup costs — all prompted, he notes, by DeepSeek releasing 'an incredible model' days earlier and proving the point.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
