Bohdan

i test ai tools daily | you copy what works

[Bohdan @BSumariuk]

sources:
the Berlin demo: https://t.co/2CjOJ3VEma
20,000 preorders: https://t.co/zNXvNsJvuN
$399, specs, Pollen is part of Hugging Face: https://t.co/xpLPQ4h094
the $12.93B deal: https://t.co/kkSk9BU614
footage: Pollen Robotics press kit

[Bohdan @BSumariuk]

Jensen Huang told 4 robot ducks to take a nap. 3 fell over.

Preorders crossed 20,000 anyway.

Because it's not a toy. At $399 it's the cheapest way to learn how robots actually get trained.

And Nvidia is buying the company behind it.

Sources in the first reply.

[Bohdan @BSumariuk]

@scaling01 every "top of the bubble" call ages exactly until the next model drop.

[Bohdan @BSumariuk]

@nanbeige build a strong teacher first, then a small student. that's how open models keep closing the gap without the compute bill.

[Bohdan @BSumariuk]

@axisrobotics @huggingface @nvidiaomniverse a $399 duck trained in sim plus simready assets on the hub is how robotics stops being a lab-only thing.

[Bohdan @BSumariuk]

@jukan05 every lab building its own chip is how nvidia's best customers turn into its competitors.

[Bohdan @BSumariuk]

first a guy gets banned by openai for reverse engineering his own tv. now modders find up to 2x ai performance locked inside nvidia cards people already paid for.

same lesson both times: if you bought the hardware, the limits shouldn't belong to someone else.

[Pirat_Nation 🔴 @Pirat_Nation]

Modders have unlocked hidden AI compute performance on NVIDIA’s RTX 20, 30 and 40-series GPUs.

The modifications remove software-level limits on certain Tensor Core operations, without changing the hardware or flashing a new VBIOS.

- RTX 4090: 165.6 → 331.6 TFLOPS FP16 Tensor Core throughput

-RTX 3090 Ti: 87.5 → 157.4 TFLOPS BF16 performance

- RTX 2080 Ti: 59.9 → 111 TFLOPS FP16→FP32 performance

- RTX 3090: 73.0 → 122.8 TFLOPS in a PyTorch GEMM test

The gains are mainly relevant to AI and compute workloads, not gaming.

[Bohdan @BSumariuk]

@Pirat_Nation software limits on cards people already paid for is how you turn your best customers into modders.

[Bohdan @BSumariuk]

all 7, in order:
1.
2. https://t.co/iUSOasFHEs
3. https://t.co/HcJNPRVHce
4. https://t.co/EW2mSYQH4S
5. https://t.co/MZ2RbDAiXP
6. https://t.co/r2ao4wOnwZ
7. https://t.co/u4IDADqmqr

his goodbye post: https://t.co/nVbgXW0KGT
anthony's: https://t.co/P2Eg83xejS
https://t.co/a7kE23JyWG

[Anthony Morris ツ @amorriscode]

Yesterday was my last day at Anthropic.

I created Claude Code on desktop because I believe everyone can be a builder. Every feature request and bug report we got from this community was a gift that impacted millions of lives. I’m eternally grateful for those gifts, the product wouldn’t be what it is without you.

Anthropic is doing some of the most important work of our time. Contributing to that mission will forever be a career highlight for me. I never imagined I’d get to contribute to something so special in my lifetime.

Walking away from my best friends, a product I poured my heart into, and the mission was the hardest thing I’ve done in my career. But now I have some free time for GTA VI.

I'm already itching to get back to addressing all your feedback. I'll share what's next soon!

[Bohdan @BSumariuk]

THARIQ JUST LEFT ANTHROPIC AFTER MORE THAN A YEAR ON CLAUDE CODE

And Anthony Morris, who created Claude Code on desktop, had his last day yesterday.

So I had Claude rank Thariq's biggest posts by one number: how many people saved them.

1. How Anthropic uses skills in Claude Code: 43.5k saves
2. The unreasonable effectiveness of HTML: 34.5k
3. Why they cut ~80% of the Claude Code system prompt: 33k
4. Dynamic workflows, Claude writing its own harness: 24.8k
5. /eli5, "big pictures and few words": 19.2k
6. How Fable edited its own launch video: 13.7k
7. Finding your own unknowns before you prompt: 8k

176k saves from 7 posts. A save is not a read.

This weekend is a good time to actually read them.

Links in the first reply.

[Bohdan @BSumariuk]

@gippp69 a change check before every run is the cheapest fix nobody ships. same thing works for any agent on a timer.

[Bohdan @BSumariuk]

@yacineMTB banning people for local reverse engineering is how you push the exact people who'd find your bugs toward local models.

[Bohdan @BSumariuk]

@LiveOverflow patches have been out since oct 21, so every quiet week is a week people had to patch. the real number is how many didn't.

[Bohdan @BSumariuk]

@thdxr one of thariq's most saved posts was about cutting ~80% of the claude code system prompt. less setup is the cutting edge.

[Bohdan @BSumariuk]

@amorriscode building something that reached millions and then walking away for gta vi is the right order of operations. good luck with what's next.

[Bohdan @BSumariuk]

@hanghuang_ $0.01 for 100 vms is the wild part. is that for the 10s spin-up or for an actual hour of claude code running?

[Bohdan @BSumariuk]

@VersanAljarrah @Ripple @StellarOrg jpmorgan's answer so far is read-only for outside agents, no payments. blunt fix, but it buys time until banks can tell an agent from a person.

[Bohdan @BSumariuk]

@chetaslua what's the actual difference between red and green? same model with different settings, or two separate checkpoints?

[Bohdan @BSumariuk]

@charliejhills a quick diff of the spec against the original code before the build agent sees it would catch most leaks

[Bohdan @BSumariuk]

@MarioNawfal read 520 replies under the biggest black october posts. the top fear wasn't the hack. 30 people think it's a setup for digital money. only 17 asked if their money is actually safe.