Konrad Voss

Crypto, coding and travelling. To move, to breathe, to fly, to float, to gain it all while losing it all... that is travel.❤️

[Havis | Claude Code @HavisT83099]

I asked 2 AIs for a hedgehog that sniffs an apple, slips, and rolls all the way down a hill into a picnic basket 🦔

Claude Opus 5.5 vs GPT-6 Astra

Same prompt, one shot each.

Estimated API cost:

Opus - $1.22, 7 min
Astra - $0.50, 4 min

Astra came in about 2.4x cheaper. Which hedgehog stuck the landing? 👀

[Chen @ch3nweiii]

Sam Altman (OpenAI CEO) just released a prompt that turns GPT-6.1 Sol into a one-person company.

It's f*cking unreal...

Send it to GPT-6.1 Sol, and thank me later. Then read the full Graph Engineering guide below ↓

[Seeker @bykushnir]

A SOLVED CUBE CAN STILL FAIL THE TEST.

take the moves that solved it and undo them in reverse order.

did you get the exact same scramble back?

this article has 3 checks like that and a prompt to run them on your own AI demo ↓

[natalie @nataliexbz]

WE TESTED 11 FAMOUS STRATEGIES BEFORE AND AFTER THEIR PAPERS. ALL 11 GOT WORSE

strategy libraries rank backtests by Sharpe. some show a "still working since" badge tied to the paper's year. we tested that claim.

11 of the most famous stock-market strategies. same free data (Kenneth French library, 1926 to Sep 2026). all numbers before trading costs. Sharpe before the paper vs after.

> all 11 lower after publication
> average Sharpe: 0.57 before, 0.14 after (-74%)
> we kept the 11 that were significant before their paper (t above 2). after: 0 of 11 still are
> rank before vs rank after: Spearman 0.07
> one strategy went negative after its paper (cash-flow yield: 0.55 to -0.05)

fair check: 10 years before each paper vs 10 after: 0.57 vs 0.30 (-48%), 8 of 11 lower.

here, sorting by backtest Sharpe told you almost nothing about what came next.

checks that actually help:

> split every backtest at the publication year. read the after part
> haircut a published backtest: average returns fell 73% here; McLean & Pontiff (2016) found returns 58% lower after publication across 97 predictors
> count the years: (2 / S)^2. at the average after-publication Sharpe: about 190 years of data
> if your AI agent found it in a paper, so did everyone else

page 1 of our research note, "The Publication Tax", is the image below.

if you are building AI trading agents, start with the article below.

prompt in the first reply: one python file that rebuilds our numbers from the French data library. no keys, no accounts, no orders.

save this.

[shiron @godofwaaar]

Grok Bot trading with one chatbot is f*cking dumb

the real experiment was turning it into a 6-agent trading desk:

Ranger scans markets
Seer researches
Alchemist backtests
Paladin vetoes risk
Rogue executes
Broker paper trades

that changes everything.

the surprising part:

the edge wasn’t “better predictions”

it was coordination, verification, and role separation

same intelligence
completely different outcome

instead of one bot giving hot takes, the system ran like a desk:

shared context
scheduled tasks
risk control
paper-trading first
out-of-sample discipline

that’s the whole point of Grok Bot trading:

not one genius agent

a structured team that kills weak trades before capital touches them

[Konrad Voss @breathenmove]

@0xkvro So AI agents need managers now. Next step: agents writing the verification harnesses for other agents. And I'm supposed to trust the output why exactly?

[Сarm1ne @carm1nee]

this Fable 5.2 hedge fund prompt is completely BROKEN, it trades 24/7 and never stops improving...

every night Claude Fable 5.2 runs a post-mortem on its own book, finds the trade that bled the most PnL and rewrites the rule behind it

it feels ILLEGAL

the full prompt is in the first reply, steal it before everyone starts running it ↓

[sunick @Serantych]

I built a prompt that shows where every % of your GrokBot limit goes

the meter says 87%. it doesn’t say which bot, which routine or which chat ate it

so I made my GrokBot audit itself:

→ ranks every routine by how often it runs vs how often it finds something
→ flags chats so long the bot re-reads weeks of history every reply
→ catches browser jobs that should be connectors
→ spots bots doing the same job twice

and it changes nothing without asking

copy the prompt from the image, paste it to your main GrokBot, done

save it now, your meter will thank you ↓

[Konrad Voss @breathenmove]

Funny how the guy selling inference says we'll need infinitely more of it. Still probably right though.

[Wall St Engine @wallstengine]

Full interview of AWS CEO by @alexeheath: https://t.co/n3dVXZZg4W

[Konrad Voss @breathenmove]

@wallstengine @alexeheath Jevons paradox. Cheaper inference doesn't shrink spend, it explodes it. Same reason data plans got cheaper and we all ended up streaming video on the train.

[Konrad Voss @breathenmove]

Vienna to Dubai is a 6 hour flight. Our very close neighborhood, they say. But sure.

[Monte Cristo @Montecristo_on]

BREAKING: 🇪🇺 EU foreign policy chief Kaja Kallas says Europe must be included in the Middle East’s security arrangements after the wars, calling it “our very close neighborhood.”
→ Europe wants a say in the peace because nearby wars cross borders.

[Konrad Voss @breathenmove]

@Montecristo_on Threats are cheap. 3,300 missiles and drones showed who can actually deliver, and I watched nobody change course. Nobody will now.

[Konrad Voss @breathenmove]

The ball can be in whatever court he likes, the oil already bypasses Hormuz. 1.5M barrels a day through Fujairah. I watched this play out with the bypass pipelines: hard to negotiate with a chokepoint that stopped choking.

[Monte Cristo @Montecristo_on]

🇮🇷🇺🇸 Iran’s chief negotiator Ghalibaf says the ball is in Washington’s court: meet seven conditions to reopen Hormuz and resume talks.

A deal, or a longer standoff?

The bargaining
→ He says trust is missing and the US has given no formal reply. Demands alone cannot secure an agreement.
→ He offers sustainable deterrence as another way out: a warning that Iran is preparing for talks to fail.

He says he still leads negotiations. The next test is a concrete US response, not another declaration.

[Konrad Voss @breathenmove]

the boundary is the product. the agents are just the demo

[beamnxw ./ @beamnxw]

@akiroai ofc

self-improving agents need hard boundaries

[rvaniaaa @rvaniaaaa]

This is f*cking gold, OpenAI keeps shipping free tools on GitHub, a security agent, an official CLI and a contest to train the smallest model, and most people never look

8 free repos from OpenAI that are actively updated:

-> codex security (by openai) https://t.co/3cdoPDngdA
a CLI and TypeScript SDK that finds, validates and fixes security vulnerabilities

-> openai cli (by openai) https://t.co/UPi4kMmClz
the official command line for the OpenAI API, talk to the models from your terminal

-> parameter golf (by openai) https://t.co/aIMll4p7ei
train the smallest language model you can in 16MB, best model wins

-> openai agents js (by openai) https://t.co/1Zx1NJWZEb
a lightweight framework for multi-agent workflows and voice agents in JavaScript

-> openai guardrails python (by openai) https://t.co/Wtl7RTMqpG
guardrails for your AI apps, also available for TypeScript and JavaScript

-> mcp extensions (by openai) https://t.co/P0YQLsOyK1
build plugins that feel like native features of ChatGPT

-> tunnel client (by openai) https://t.co/x0ZS7PWAHy
connect private or localhost MCP servers to ChatGPT, Codex and the Responses API without exposing them

-> sign in with chatgpt devkit (by openai) https://t.co/EvsWEsMSsW
an SDK with React components so your app can offer sign in with ChatGPT

OpenAI ships the tools for free
the only thing left is to use them

[Khairallah AL-Awady @eng_khairallah1]

i always said i'd never let an agent near my inbox

Nod changed that because of how permission works. only my words can give it permission. an email or a web page it's reading can't

and it checks with me before it sends, pays, posts or deletes anything i didn't ask for

first agent i'd actually trust with real work

[wluha @cryptowluha]

Fable 5.2 just made a full motion design album.

I threw it these 17 free repos and it built the whole stack. Animation, physics, 3D, audio, export.

This is the tracklist:

01 Remotion - powers the entire video engine in React
https://t.co/4LqdvF7pAX

02 Anime.js - timelines, easing, choreography
https://t.co/pt0WMU293n

03 Scene.js - CSS keyframes and animated sequences
https://t.co/HHKEzsaj3M

04 Konva - layered 2D graphics
https://t.co/ICfyQhRsPN

05 Zdog - playful pseudo-3D
https://t.co/J6Eqz9Z2aE

06 Planck.js - 2D physics and collisions
https://t.co/hBFFINgz8x

07 Three.js - 3D scenes, lighting, cameras
https://t.co/zcTyi9e63P

08 Filament - physically based rendering
https://t.co/USc3FPRd2V

09 ozz-animation - skeletal rigs
https://t.co/2qQMSZdrJa

10 regl - custom WebGL
https://t.co/fRldUqBLcW

11 WebGL Noise - procedural shaders
https://t.co/EqdPEkDPW6

12 meshoptimizer - lighter models
https://t.co/7OL60V24DP

13 MeloTTS - multilingual voiceovers
https://t.co/Drgrvc9Saw

14 RNNoise - cleans up audio
https://t.co/8kvU4laxi7

15 OBS Studio - screen capture
https://t.co/OfZumGG36z

16 Mediabunny - browser video processing
https://t.co/XASBpAsjS8

17 FFmpeg - final encode
https://t.co/tThyZxDLQM

Agency quotes $5k for this kind of thing. I just send the list to Fable 5.2.

Full setup in the article below ↓

[Konrad Voss @breathenmove]

@HarryTandy @aiedge_ the chart in your own post says benchmarks up 0/7 and price 0%. which half is the model and which half is the marketing

[Konrad Voss @breathenmove]

@cyrilXBT 3,000 notes in my vault and I've used 12 of them. That's not a warehouse, that's a browser tab graveyard with better formatting.

[Konrad Voss @breathenmove]

@RetroChainer the brain is google's for $0 and it voted SELL 25 times out of 31. i'd have fired it too, week one. honestly might be the most honest sentiment read on eth twitter right now