Veliorix

The future called. I picked up.

[Veliorix @veliorix]

@Newtonixe two years ago a short like this needed a small studio. now it needs one person who has something to say. the tools caught up, the stories are the hard part again

[Veliorix @veliorix]

@Lumesssss what makes this hard is that the first steps won't look dramatic. AI already writes a big chunk of the code that trains the next model. the line gets crossed gradually, not on one decision day

[Veliorix @veliorix]

@Lumesssss the scariest part isn't the 2027 to 2030 window, it's the review speed. if evaluation stays human-paced while models ship in weeks, the safety check turns into a formality. the evals have to get automated too

[Veliorix @veliorix]

One year ago, the Fed Chair was asked what happens if AI investment SUDDENLY CONTRACTS. This week the market answered.

October 29, 2025. A Politico reporter asks Jerome Powell how worried he is about a sudden contraction in tech investment, and what lessons he takes from the 1990s.

His answer: this time is different. The dot-com names were "ideas rather than companies." Today's leaders actually have earnings, business models and profits. And the data center buildout is one of the big sources of growth in the economy.

November 5, 2026. Nvidia closes down 9.3%. AMD down 17.3%. Broadcom down 10.4%. The biggest tech selloff since Covid, after the midterms brought candidates and ballot measures that want to slow new data centers.

Powell was right about the earnings. Nothing on Thursday was about AI companies losing money. What got sold was the buildout, the exact thing he called one of the biggest sources of growth.

So two things are true at once:

- the companies are real
- the bet on an endless data center buildout just got repriced

That is a different story from 2000. Back then the products barely existed. Now they work and keep getting cheaper to use. The open question is how much concrete and silicon the world needs to pour under them.

His full answer is in the video.

[Veliorix @veliorix]

@trirnolo and the second trade is the one most people can actually make. you don't need a brokerage account to build on cheaper tokens

[Veliorix @veliorix]

NVIDIA JUST LOST OVER $400 BILLION IN ONE DAY. THE PRICE OF USING AI DIDN'T MOVE A CENT.

Wall Street just had its biggest tech selloff since Covid. Nvidia closed down 9.3%. Broadcom down 10.4%. AMD down 17.3%. Nvidia is now at its lowest level since April.

No earnings miss caused this. After the midterms, voters backed candidates and ballot measures that want to slow down new data centers, and investors started pricing in fewer GPU orders.

Look at what actually got sold:

- chipmakers
- memory makers
- the neoclouds that rent out GPUs
- anyone whose revenue depends on building the next data center

Now look at what nobody sold. The cost of running a model.

This week Revolut showed its 2,000 engineers running on an open model at about $0.30 per million tokens, roughly a tenth of what it paid Anthropic and OpenAI. Nothing about today's crash changes that number.

Wall Street is repricing the companies selling shovels for the AI gold rush. The people actually using AI had a normal Thursday: same models, same prices, same output.

If you build with AI, your stack didn't get worse today. Your API bill didn't go up. The only thing that changed is how much investors are willing to pay for the hardware underneath it.

[Veliorix @veliorix]

@Flikore the part every crash headline skips. and your graph not a line post earlier was a great read, the "check that can fail" bit stuck with me

[Veliorix @veliorix]

@AIDuckLabs Wall Street sold the shovels, not the gold

[Veliorix @veliorix]

@Flikore 31 to 17 minutes with the same tokens is the cleanest case for drawing agents as a graph instead of a line

[Veliorix @veliorix]

@fluid_ai_ the hard part of shared memory is forgetting. if every tool can write to it, it also collects every wrong assumption and outdated decision. curious how Living Brain decides what's stale before it retrieves it

[Veliorix @veliorix]

@artembkk the consistent character part is the bigger deal for me. at 4 billion images a week the feed fills up with polished but generic visuals, so a recognizable style becomes the moat, not the ability to make an image at all

[Veliorix @veliorix]

@Grirku Opus as an on call reviewer instead of the main worker lines up with that Alibaba harness paper going around: an independent critic in a fresh context added 5.9 to 9.5 pp on Terminal-Bench for about 1.2x the cost. the "test fails twice" trigger is the smartest part of this tree

[Veliorix @veliorix]

@f0rgent yep. the catch most summaries skip: it runs on 56 Nvidia B200s, and frontier models still handle UI and vision. cheap by default, expensive by exception

[Veliorix @veliorix]

REVOLUT CUT ITS AI CODING BILL BY 90%. YOU CAN COPY THE IDEA WITHOUT 56 GPUS

Revolut moved about 2,000 engineers off Claude Code and Codex as the default and onto GLM, an open-weight model it hosts on 56 Nvidia B200s. About $0.30 per million tokens. Roughly a tenth of what they paid before.

You probably can't host a 380GB model. You don't need to. The logic works at small scale too, and Revolut's infra lead says open source is now a company priority for exactly this reason.

how I'd set it up:

- keep Claude Code. your team already knows it
- point it at an open model through an Anthropic-compatible endpoint. Zhuip AI and OpenRouter both have one, it's a base URL and a model name in settings
- send the routine stuff there: tests, refactors, small tickets, code summaries
- keep Opus for architecture, nasty bugs and anything visual. Revolut does the same with frontier models for UI and vision
- stop rationing tokens. Revolut's engineers had strict budgets before the switch, and honestly that part sounds like the real bottleneck
- try one repo for a week first
- track cost per finished task, retries included. price per token lies

before: every edit billed at frontier rates
after: frontier prices only where they earn it

config in the reply

[Veliorix @veliorix]

@deeponailabs agents check their own edits less than 14% of the time, and a cheap critic in a fresh context fixes a big chunk of that for about 1.2x the cost. that's the most underpriced upgrade in agent design right now

[Veliorix @veliorix]

@LendaHere metered tokens make people careful, flat cost makes them curious. Revolut's engineers went straight to debugging and full Jira tickets once the caps were gone

[Veliorix @veliorix]

@vglonix the WiFi sensing part is real, CMU did it in 2022 with $30 routers. the second shape is what keeps me up

[Veliorix @veliorix]

@LendaHere the blackjack bought a week. the $11M raise a few weeks later saved the company. nobody retells the boring part

[Veliorix @veliorix]

@deeponailabs that's the catch for small teams. Revolut keeps 56 B200s busy with 2,000 engineers. a 10 person team can't, which is why the setup above runs the open model through an API instead of owning GPUs. cheaper tokens, no idle hardware

[Veliorix @veliorix]

@vglonix exactly, and it cuts both ways. if the open model nails tests and refactors on the first try, you get the same result for a fraction of the cost. the one repo week is how you find out which side of that line your codebase is on