Velmonta

AI, taken apart.
3D breakdowns of the chips, models and robots everyone's talking about.

[Velmonta @velmonta]

Show Claude or GPT-5.4 a picture of a red heart and it writes different code. 20 out of 20 times.

That's from a new Google DeepMind paper, published Oct 27. What they did is simple.

They gave three top models a basic coding task: go through every item in a tree-shaped data structure. There are two normal ways to write it. A loop, or recursion, where a function keeps calling itself. With no image, Claude Opus 4.7 and GPT-5.4 picked the loop every single time.

Then they added a small picture and asked the model to look at it. With a red heart, both switched to recursion in 100% of tries. For Gemini 3.1 Pro the trigger was a yellow checkmark, and recursion went from 40% to 90%.

They tested 288 simple symbols, 12 colors times 24 shapes. Random noise and normal photos did nothing. Writing the words "a red heart" instead of showing it barely did anything either. So it's something in the pixels themselves.

It also worked when the image was slipped in with no instruction at all. Some symbols still flipped Claude every time.

The weird part is the explanations. When the image changed Gemini's choice, its reasoning gave the image credit only 25% of the time. The rest of the time it explained the switch with normal sounding technical reasons.

Why this matters: agents now read screenshots and click around websites. A random icon on a page could nudge what they do, and the reasoning log might not show it.

Two caveats. It's 20 runs per setup on a toy task, and the authors call it preliminary. They frame it as a transparency problem, not a hack.

Which icon would you test first?

[Roan @RohOnChain]

if you want to build an AI trading system from scratch, start with this article

you will end up with real market data, real backtests and a bot that trades for you 24/7 https://t.co/7bOgT1HCZf

[Velmonta @velmonta]

@volero87 you are who you live with. even if ypu're claude

[Velmonta @velmonta]

@HavisT83099 all three agreeing on the crash is the real benchmark

[Velmonta @velmonta]

@nightraccai checking is the only ai edge nobody screenshots

[Velmonta @velmonta]

One AI prompt uses about 5 drops of water. Google finally measured it.

For a typical Gemini text prompt it's 0.24 Wh of electricity, 0.26 ml of water and 0.03 g of CO2. Google says that's about the same as watching TV for less than 9 seconds.

So where does that energy actually go when you hit send?

Most of it, 58%, goes to the AI chips doing the "thinking", the math that picks your answer one token at a time. Another 25% goes to the regular CPU and memory that keep those chips fed. 10% is spare machines kept running on standby so the app doesn't lag when everyone logs on at once. The last 8% is the building itself: cooling, and turning grid power into something the racks can use.

The water is mostly cooling. Chips turn electricity into heat, and lots of data centers get rid of that heat by evaporating water.

OpenAI's number is close. Sam Altman put an average ChatGPT query at 0.34 Wh, about what an oven uses in a little over one second.

And it's dropping fast. Google says energy per prompt fell 33x in a single year, mostly thanks to better software and chips.

The catch: these are typical plain text prompts. Long reasoning, image and video generation and agents running for hours use a lot more. Training the model isn't counted at all. And small numbers add up. ChatGPT alone was handling around 2.5 billion prompts a day in mid 2025.

So one prompt costs almost nothing. Billions of them a day are why Microsoft, Google and Amazon are signing deals with nuclear plants.

Would you have guessed higher or lower?

[Roan @RohOnChain]

if you want to build an AI trading system from scratch, start with this article

you will end up with real market data, real backtests and a bot that trades for you 24/7 https://t.co/7bOgT1HCZf

[Velmonta @velmonta]

@bottegamd everyone wants the agent that earns for them. nobodys asking who audits it

[Velmonta @velmonta]

@ch3nweiii does it actually finish or just finish confidently? curious if anyone ran it on a messy repo

[Velmonta @velmonta]

@rosenacry real question for anyone hiring juniors now: ore u hiring fewer or hiring them to run the agents?

[Velmonta @velmonta]

There are 4 Google AI chips orbiting Earth right now. Up there, a solar panel makes up to 8x more power than the same panel on your roof.

That's the whole bet. AI is hitting power limits on the ground, and in the right orbit the sun almost never sets.

The satellite went up on October 1. Last week a TPU on board ran Gemma 4 and sent back a haiku while passing over Norway: "Blue marble below / Data streams in silent grace / Starlight feeds my core."

The weak spot is memory. In Google's radiation tests the HBM started acting up first, but only after about 3x the dose a shielded chip would get in 5 years.

Next up: two linked satellites in 2027. Google's own math says it gets close to the cost of a ground data center if launches drop under $200/kg, maybe by the mid-2030s.

Would you trust a data center you can't walk into?

[Addy Osmani @addyosmani]

https://t.co/2gyp2Erpxx

[Velmonta @velmonta]

@LRudL_ late 2026 market correction is the only one left to grade. eight weeks to go👀

[Velmonta @velmonta]

@ch3nweiii the check step is the one everyone skips. without it the agent can "fix" the same test 40 times and call it progress

[Velmonta @velmonta]

@godsystemsmd real question whats one agent you'd actually pay for today? mine is still the one that writes the boring tests

[Velmonta @velmonta]

@0xmurasame both of them saying 2026 or 2027 out loud is the part i'd bookmakr. easy to check who was right in 14 months

[Velmonta @velmonta]

DeepSeek is putting at least 160,000 of these in one data center in Inner Mongolia. Not one of them is Nvidia.

This is Huawei's Ascend 950DT, taken apart. 144 GB of memory Huawei makes itself, about 4 TB/s of bandwidth, 1 PFLOPS at FP8 on paper.

Now do the math. 160,000 of them is around 23 petabytes of fast memory and 160 exaflops of FP8 in one place. Each chip is a lot weaker than Nvidia's best. China's plan is just more of them.

Yesterday Chinese media reported DeepSeek wants to train V5 on these, not just run it. This is the lab that knocked 17% off Nvidia's stock in a single day in January 2025.

Who ships the better model in 2027: xAI with 860,000 Nvidia GPUs, or DeepSeek with 160,000 of these?

[Addy Osmani @addyosmani]

https://t.co/2gyp2Erpxx

[Velmonta @velmonta]

Amazon wants delivery drones in nearly 500 US cities by January. Last week they went live in Idaho. Tallahassee is next.

So here's roughly what'll be flying over your roof, taken apart.

Most of it is a battery and 4 motors. The part that matters is the little board in the middle. It corrects all 4 motors about 1,000 times a second, so a gust doesn't flip it.

The AI lives in the camera under the nose. There's no pilot with a joystick, so it has to spot wires, birds and people on its own.

I built the whole thing with Claude Code. Prompt's in the reply.

[Roan @RohOnChain]

if you want to build an AI trading system from scratch, start with this article

you will end up with real market data, real backtests and a bot that trades for you 24/7 https://t.co/7bOgT1HCZf

[Velmonta @velmonta]

xAI is putting 860,000 of these in one building.

This is one of them, taken apart. Liquid cooling on top, two compute dies in the middle, 8 stacks of memory around them. 208 billion transistors in one chip.

Now picture 860,000 of them running at once.

What should I take apart next?

[Roan @RohOnChain]

if you want to build an AI trading system from scratch, start with this article

you will end up with real market data, real backtests and a bot that trades for you 24/7 https://t.co/7bOgT1HCZf

[Velmonta @velmonta]

the math is clean but the number that's missing is what keeps the five agents running after the first hour. 96 agent-minutes of Astra at xhigh is roughly $30-40 in tokens. do that daily and you're spending $1K/month on agents before the business has a single customer. the architecture is real. the unit economics start after the build.

[Velmonta @velmonta]

the interesting part isn't two brains. it's what happens to the cost curve when you route cheap thinking to one model and expensive thinking to another. Astra bills $50/M output tokens. Kimi doesn't. putting research on the cheaper brain and execution on the expensive one is the first architecture that actually treats token spend as a design decision.

[Velmonta @velmonta]

This is the most underrated AI dataset on the internet right now

12.3 GB of real production trading data
not simulated. not backtested. real.

79 anonymized features straight out of Jane Street's systems
4.5 million rows in training alone
scored against six months of future data
that didn't exist when submissions closed

read that again.

the model had to predict instruments it had never seen
on data that hadn't happened yet.

every winning notebook is still free.
the evaluation API is still live.

if you're using AI anywhere near markets
and you haven't gone through this,
you're training on the wrong data.

[Rossst.03 @Rossst_03]

Jane Street published 12.3 GB of its own production trading data and paid $50,000 to whoever forecast it best.

$120,000 across the top ten. 3,757 teams tried.

the data, the evaluation API and every winning notebook are still up, still free. what is actually in there:

1. 79 anonymized features and 9 responders, one row per instrument per timestamp, straight out of their production systems

2. you predict responder_6. scoring is a sample weighted R squared taken about zero, so predicting nothing scores exactly 0 and a bad model goes negative

3. the training test set alone is about 4.5 million rows

4. new symbol ids show up in the future test set, so the model has to handle instruments it has never seen

5. submissions froze on 13 January 2025, then got scored for six more months against market data that did not exist yet

6. 8 hour notebook limit, 9 in the forecasting phase. not the best model, the best model that finishes

the prize is gone and the board is settled. what is worth taking is the evaluation design. they made the answer depend on six months nobody had.

[Velmonta @velmonta]

Klarna replaced 700 employees with AI.

2.3 million conversations in 30 days.
Resolution time: 11 minutes → under 2.
$40 million in projected savings.

14 months later the CEO said three words:

"We went too far."

Here's what everyone missed.

Headcount: 5,527 → 3,422.
That part never reversed.

The humans they rehired came back as freelancers.
No benefits. No guaranteed hours.
The CEO called it "Uber-type" support.

AI didn't replace the workers.
It gave the company a reason to reprice them.

Full-time payroll became variable spend.
The kind that disappears in a bad quarter.

If your company is "exploring AI" right now,
this is the playbook they're reading.

Not whether AI works.
How AI changes what your role costs.

[rari🈀 @0xwhrrari]

https://t.co/GoJ0DWB7Ve