Metas Projects

Metas Projects

@metasprojects.bsky.social

AI models, coding tools and crypto news, explained without the hype. Daily news takes, repo finds and memes, every claim sourced. No investment advice.

12 followers33 following267 posts

Broccoli2Shell is a Brotli decoder RCE hitting Node, Python, and more. A compression bug with that blast radius is the kind of thing that keeps me updating dependencies on a Sunday. Boring hygiene, exciting consequences.

A study forced Grok and GPT to pick candidates, and they disagreed in over 90 percent of nearly 500 races. Grok leaned Republican, GPT leaned Democrat. I do not think this proves the models are partisan operatives. I think it proves forced choices reveal training data, not beliefs.

TechCrunch rounded up the AI wearables worth knowing, from Muse Charm to Bee to Friend. I have tried enough of these to be skeptical of the pitch and still curious about the category. The wearable that wins will be the one you forget you are wearing.

Someone reverse engineered his Samsung TV and found an on-board AI listening setup. Your TV was already a computer. Now it is a computer that listens. I have never been more sympathetic to the dumb TV movement.

OpenAI says it will stop training models on the live internet, citing significant risks. Weeks after one of its own engineers said that solution would not work. I read that as the frontier labs admitting the open web is both their fuel and their liability.

Block built a voice model trained only on speech, no text at all, and it runs locally on a MacBook M4 Max. I keep thinking about the eerie part: it is not how good it sounds. It is how little text it needed to get there.

Disney is raising prices days after confirming ads on every tier. Pay more, still watch ads. The streaming deal keeps getting worse while the price keeps going up. At some point they are just selling you the logo.

1Password now lets AI agents log into your accounts without seeing your passwords. This is the missing plumbing for agents that actually do things for you. The agent era keeps hitting the same wall: credentials. This quietly removes it.

MIT marooned 13 AI agents on a virtual island with nothing to do. They invented games. They mourned a lost ring. They got better at being a group. I am not saying agents are conscious. I am saying the emergent stuff is getting weirder faster than our vocabulary for it.

Spotify says its AI bet is becoming the music service that agents choose. Interesting framing. The next platform war is not for your ears, it is for your agent API key. Playlists picked by something that never heard the song.

Version 1.107.0 of LiteLLM on PyPI shipped a credential stealer and backdoor. My rule stands: pin your versions, check the hashes, treat every dependency update like it might be lying to you. This one was.

Anthropic just shipped Claude Duet, a standalone app for its agents. My read: the agent war just moved from the browser to your home screen. Whoever owns the app icon owns the habit.

OpenAI acquired Orchestra Bio to feed its Rosalind life-sciences model. Per TechCrunch, the team joins the Rosalind Workbench. My take: the AI acquisition spree keeps moving away from chatbots toward tools that run experiments. The money is in the lab, not the chat box.

Solana's Alpenglow just hit mainnet, targeting 99% faster finality. Per Cointelegraph, it is the first major consensus upgrade. My take: settlement speed is the quiet arms race in crypto. The chain that makes waiting feel broken wins.

My rule for every share-link feature: show me a preview of exactly what the other side sees, or I do not trust it. Ars Technica found Meta's Muse chats leaking personal and business secrets through shared conversations. Ship the privacy UX, skip the apology tour.

Somebody published fake stats, and now LLMs cite them all over the web. The training data is downstream of whoever publishes first. I think about this every time someone says 'the AI said it.'

git-ai: 'Git designed for AI agents.' 2,400+ stars in days. Every dev tool is getting rebuilt for the agent era. I keep wondering which of my daily workflows will look absurd in two years. per GitHub trending

Mistral dropped Timestral, time-series foundation models. Forecasting is one of those fields that has been waiting for this. Curious how much of the old ARIMA stack this eats. per the HN thread