Giorgio Zinetti @giozzi

We are fun

Stani @StaniKulechov

2 weeks ago

Working with three Italians on the team has been the most fun I’ve had in a while. 🇮🇹

122 21 762 68K 29

Considering starting a climbing club for startup folks around Zürich. Let me know if interested and RT for exposure.

@mewwts Come to mito ❤️

@lalazar Probably not a website

@mewwts @Padierfind Btw used to climbed 7c, now I’m out of shape

@mewwts @Padierfind Thanks 🙏

@mrosberghaus no.. proper code :D

When did you write your last line of code?

Champions league

Wonderful @wonderful_ai

4 weeks ago

FC Bayern as a club never settles for anything less than world-class. We are proud to share that Wonderful is now an Official Enterprise AI Partner of @FCBayern.

28 35 389 246K 31

𝐅𝐂 𝐁𝐚𝐲𝐞𝐫𝐧 𝐱 𝐖𝐨𝐧𝐝𝐞𝐫𝐟𝐮𝐥 Der FC Bayern und @wonderful_ai haben eine dreijährige globale Partnerschaft geschlossen. Als offizieller Enterprise-KI-Partner des FC Bayern wird Wonderful den Verein dabei unterstützen, KI in ausgewählten Bereichen des Vereinsbetriebs in die Praxis einzuführen, angefangen bei ausgewählten Fan-Services. ℹ Mehr Infos: fcbayern.com/RtKnE

@lalazar ❤️

@lalazar LFG

Wonderful has raised a $550M Series C at a $5 billion valuation. This latest funding was led by @insightpartners, with @salesforce joining as new investors and @IndexVentures, @BessemerVP, @IVP, @VineVenturesLP, and @9yardscapital returning. In 20 months since launch, we're in production with hundreds of agents, systems, and workflows across a dozen verticals in over 30 markets. With this new capital, we’ll be overdelivering for our customers, hiring hundreds of cracked engineers, FDEs, and former founders, and expanding our product capabilities across the entire enterprise. Stay tuned 👀

After 2.5 years, I’ll be stepping down as CTO of the Cardano Foundation at the end of August. Time really flew. I’m proud of what we built, but even more grateful for the people I got to work with along the way. First, a few weeks off to welcome our daughter ❤️ Then a new adventure starts in September. More soon :)

The “For You” tab on X has become a collection of garbage. Every other post is manufactured outrage from someone pretending to be an idiot because it pays better than having an actual opinion.

agi is coming. the upside is absurd — curing cancer, clean energy, post-scarcity. demis dropped his framework today: us-led standards body, finra-style, tests every frontier model on earth before release. my instinct: why us-led? we just watched washington freeze anthropic's models overnight with zero playbook. but every alternative is worse. labs self-regulating grade their own homework. international bodies are too slow. europe has no models to lead with. the most important technology in history, and our best idea is "wall street's regulator, but for superintelligence." i don't have a better answer. does anyone?

Demis Hassabis @demishassabis

3 months ago

x.com/i/article/2076…

2K 5K 24K 15.8M 31K

Doctors are obsolete

Karan Singhal @thekaransinghal

3 months ago

♥️ GPT-5.6 is a major step forward for health, both at the frontier and at cost. These models push the frontier of performance per dollar, bringing the best health intelligence to all. The smallest variant, GPT-5.6 Luna, evaluated at the lowest reasoning effort, outperforms

57 128 958 1.4M 329

The RL framework behind GLM-5.2 is fully open source. The full post-training of GLM-5.2 ran on it in about two days. The same stack sits behind the entire GLM series, from 4.5 to 5.1. It is called slime, and it is built around one idea. Keep a single RL kernel, and push all the variety into data generation. Let me explain what that means. Every RL run has two halves. One generates experience, where the model produces responses and something scores them. The other learns from it by updating weights. The learning half is mechanical. It reads samples, computes a loss, and steps the optimizer, the same way whether the model solves equations or drives a browser. What changes between tasks is generation. A math run answers in a single turn and grades the result. An agent run loops through tool calls, reads results, and only then earns a reward. slime draws the line right there. The learning half stays fixed as one kernel, and everything that differs becomes a new way to generate data. Under the hood, it wires Megatron for training to SGLang for rollout, with a Data Buffer between them that owns prompts, custom data, and generation. Most RL stacks grow into a pile of disconnected trainers, rollout services, and agent frameworks. slime refuses that. Multi-turn tool use, sandbox interaction, environment feedback, and verifier rewards all enter as data generation, not as forks of the loop. So an agentic workload runs on the same loop a math run uses, and the kernel never changes. A few things follow. → It is battle-tested. The loop is validated by shipping real GLM models, and it also supports Qwen3, DeepSeek V3, and Llama 3. → Correctness comes first. RL bugs are silent, so slime keeps the dataflow explicit and treats CI, reproducibility, and fault tolerance as real engineering. The proof is the ecosystem on top of it. Dressage, Miles, vime, Relax, OpenClaw-RL, P1, and TritonForge all build on slime without touching the core loop. The lesson is not that RL needs a bigger framework. It is that the variety belongs in data generation, and the training loop should stay small enough to trust. GitHub repo: github.com/THUDM/slime (don't forget to star 🌟) Since we're talking about RL, I wrote a full breakdown on fine-tuning LLMs with RL in 2026. Including how to skip manual reward engineering with automatic LLM-graded rewards. The article is quoted below.

Akshay 🚀 @akshay_pachaar

7 months ago

x.com/i/article/2029…

31 285 2K 921K 4K

🚨SecondFi: more than $20M stolen, or taken hostage - When the nonce is predictable, the key is public. Last Tuesday, SecondFi wallets were drained at scale. Users were doing nothing exotic, just signing transactions like any other day. Then the funds were gone. ▶️What happened. Two waves. First, external attackers drained wallets: ~16M ADA plus NFTs, around $2.5M. Then a much larger move: ~129M ADA, around $20M, swept by SecondFi itself to a third party custodian, "to keep it safe." Users were robbed by attackers, and users were swept by the platform!! ▶️The bug. Every signature scheme in this family uses a per signature secret number, the nonce. The one rule: it must be secret and unpredictable. EdDSA (Ed25519, what Cardano uses) makes it deterministic on purpose, derived from the private key and the message together. Deterministic, but secret. The secret input is the whole point. SecondFi's implementation derived it from public data alone. The nonce stopped being secret. Public in, public out. ▶️Why it is fatal. Call it what it is: worse than nonce reuse. Reuse needs two signatures from the same key to recover it. Here the nonce can be recomputed by anyone from data already on chain, so a single signature is enough to reconstruct the private key. Every transaction a user ever signed is a public disclosure of their key. I pulled signatures off chain and rebuilt the keys to confirm. One signature, every time. The chain was not attacked. It was read. ▶️The "rescue." The second wave was the platform sweeping ~129M ADA to a custodian, framed as protection. The fix for "your keys are cryptographically exposed" was "trust us to give it back." A cryptographic failure patched with a social promise. "We will return the funds" is not "we are cryptographically incapable of keeping them." The incident did not create that gap. It made it visible. ▶️The ownership trap. Once a key is publicly reconstructible, holding it proves nothing. Attacker, custodian, original user: three parties, same cryptographic claim. So how does a real owner prove it? The elegant exit is a zero knowledge proof of knowledge of the 24 words, the seed, without revealing it. Knowledge of the mnemonic is the one thing a chain scraper does not have. Prove the preimage, not the key. ▶️ How did it get in? Honest mistake, or planted? Dropping a single secret input from the nonce is a tiny change. Exactly what a tired engineer ships, and exactly what a supply chain or insider attacker plants, because it is deniable and pays out silently to whoever reads the chain. I do not know which one this was. The commit history will tell. Cryptography does not care about good intentions. It enforces what you built. SecondFi built a scheme where the nonce was predictable, so the keys were public, so the funds were anyone's. The rest is just who read the chain first. Stay safe. -- (obv, Ledger users are not affected. Nonce and signatures are computed inside the secure element)