20VC with Harry Stebbings

Jev Creator on Will US Open Models Catch Chinese | The Cost of Intelligence & Tokens | Diogo Almeida: summary

YouTube summary9 sectionsWatch on YouTube ↗

This is an AI-generated summary of the YouTube video "Jev Creator on Will US Open Models Catch Chinese | The Cost of Intelligence & Tokens | Diogo Almeida" (20VC with Harry Stebbings), made with Samuraize and published by Samuraize. It condenses the YouTube video into 9 titled sections you can read in a couple of minutes, each linking to the moment in the video it covers.

1
Filed under💻 Technology0 comments🍱 Add to trayReport
Study this

Jev Creator on Will US Open Models Catch Chinese | The Cost of Intelligence & Tokens | Diogo Almeida

20VC with Harry Stebbings

切

Reliability as bottleneck 0:00

Reliability is framed as the biggest bottleneck in AI. The point is not uptime or throughput, but the steadiness of intelligence itself. Models can solve hard problems sometimes, yet still fail at simpler work like accounting. That high variance leaves you with Einstein most of the time and Mr. Bean the rest, and that is why automation keeps breaking down.

切

Data before compute 5:00

The path forward is said to run through the right task, then data, then compute, then algorithms. Scaling alone has already shown its limits, and data matters more than compute once you care about background software that must keep working. Good data is not just stored in company lakes. It is shaped for the task, often created from scratch, and judged by whether it moves the model forward.

切

Reliability over benchmarks 15:01

The speaker says the field wastes money when it chases hype, cheap copies, and benchmark numbers. What matters, he says, is not speed or cost but the reliability of intelligence. He argues that people miss this and that developers understood it faster because they care less about scores and more about whether the tool works on new tasks. Prompting also matters, but the main goal is to make the model smarter and more dependable.

切

Software as a new layer 26:30

He says AI should become a fourth basic logic gate for software. That would let you remove the line between what can and cannot be automated, and make software itself far more capable. He thinks enterprise AI is still mostly a thin layer on top of old products, because unreliable models cannot safely make high-stakes decisions. The result is user-facing slop, not deep change.

切

Workflow beats distribution 29:30

SAS companies are best placed not because they own distribution, but because they know the workflows and problems best. The speaker thinks automation needs real engineering and should be built around narrow tasks, like changing a TurboTax interface or twisting a bottle cap, rather than around AI coworkers that mimic humans. Owning intelligence makes sense only if you assume companies will train their own systems, which he does not.

切

Benchmarks and model truth 40:31

He says benchmarks are misleading and often gamed. Raw intelligence per token may be improving, but only a little, and mostly when a task is put in distribution, as coding was. He thinks top labs are mostly truthful, while weaker labs and many open model teams are pushed to benchmark max. Real progress, he says, comes from software around models, not from claiming proprietary intelligence.

切

Benchmarks and truth 45:01

Benchmarks matter too much, and that gives too much power to the people who run them. In an ideal world, no one would care about them. Top labs do not need to cheat to stay high up. They want to be truly good, and morale matters. The real question is whether American open source models need support against Chinese ones.

切

Security and demand 46:31

The threat depends on the use case. A model used raw is different from one trained with private company intelligence, bias checks, and guard rails. He would not use a Chinese model as a vanilla coding agent, but he thinks the field is too locked into one narrow view. He also thinks AI demand can drop fast when buyers see what it can really do, since many are already overestimating it.

切

Hiring and alignment 1:00:30

He says he chose Andreessen Horowitz and Sequoia for alignment, not money. He values their philosophy, especially Sonia Wong’s judgment, because she gives a different strategic view and helps him make great decisions slowly. He dislikes traditional business people who act on autopilot and wants to understand the why behind important choices. Hiring is hardest for the data team, where he needs people who can help build the company’s own way of making data. He wants obsessive people who love technology, are not mercenaries, and can push the boundaries of machine learning and product work.

AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.

Summarize your own YouTube video

Paste a YouTube link, article, PDF, ebook or slide deck and get a summary like this in seconds. Free to try, no sign-up needed.

⚔️ Try the YouTube summarizer

Discussion

Sign in to join the discussion. Sign in

More from the Bento Box

Browse the Bento Box →

With your OK, Microsoft Clarity and Google Analytics set cookies (data goes to the US) so we can see what breaks. Cookie details