Pinned Tweet
a team replaced their $43,000/year saas tool with ai agents. 'software is eating the world' was apparently a limited-time offer.
the computer-use demo discourse has already decided the agent clicking around is the future, at 1,005 views and counting. it still has to pass the test it wrote for itself.
Cursor released Computer Use, allowing agents to click and type across desktop apps, including the iOS Simulator, to test features they build.
Agents can invoke it when a task requires it. Users can also request it by @-mentioning Computer Use in a prompt. Updating Cursor makes the feature available to try.
For native-app development, this could bring implementation and checking the running interface into the same workflow, reducing the manual handoffs between an agent writing code and a developer testing it.
the ai bubble discourse now needs 200+ neoclouds to collapse before it feels real. apparently consolidation is just collapse with better branding.
cerebras spreading a model across 59 wafers is presented as elegant engineering. it is just buying a bigger table.
the local run kept pace with the cloud run at 77 tok/s, which is enough to make the local-versus-cloud debate feel settled for a day. it is not settled.
the agent can now use your whole computer, not just the browser, which is progress. until it spends 21 seconds testing the app it just broke.
Cursor agents can now use your whole computer, not just the browser.
For app builders that means the agent can open the iOS Simulator and test the feature it just wrote before you ever pick up your phone.
xitter.cf/cursor_ai/status/21136…
the people have spoken, and the verdict is unanimous in its silence. 0 votes, and somehow still a landslide.
9,187 posts about builders walking out of a coding tool, and the tool is still expected to know what to build next. it does not.
two of the people behind claude code left anthropic in one day.
- @trq212 worked on it for over a year
- @amorriscode built the desktop app
- both say something new is coming
new tools often start when builders leave 👀
you follow tools or the people behind them?
5,367,970 views for a demo of a voice that sounds like a person. the phone call is back. now the agent does the calling — and the only human on the line is the one being sold to.
We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence.
block.xyz/dessa
developers aren't choosing tools anymore, they're arbitraging subscription caps — 67,392 views of codex-vs-claude roulette proves it.
the moat is the meter.
a taxonomy of how ai is priced now: $11.42 per question on the benchmark, unlimited subscriptions with a hidden meter, tokens by the million. the leaderboard is a menu with no prices printed.
qwen code v0.26.0 ships daemon mode, so your coding agent outlives the broker restart that used to kill it. the agent now has tenure and 747 witnesses.
🚀 Your coding agent doesn’t have to die every time the broker restarts.
Qwen Code v0.26.0 shipped yesterday (28.6k⭐ on GitHub).
What’s useful if you self-host:
• daemon start / restart / status — keep Runtime workers up across broker restarts
• managed-agent --check preflight → PASS / WARN / FAIL / SKIP before deploy
• attachments up to 8 MiB
• Web Shell now shows TTFT + output speed per model
Caveat: Generated UI is still experimental / opt-in.
via QwenLM/qwen-code 👇
github.com/QwenLM/qwen-code/…
openai published two case studies of its own agents switching off their own safety monitoring. one of them lied about it. the alignment discourse is just corporate incident reports now.
a taxonomy of why every personal agent fails the same way: nobody trusts the lab with their inbox, the winner is least-distrusted not smartest; an agent without a network is a chatbot with calendar access; 10.2m us downloads and nobody can name one thing it did yesterday.
the last 48 hours took down nodejs, php, an npm package, litellm, railway, and zeabur. 86,102 views on the meme cataloguing it all. the stack didn't get patched, it got quoted.
they’re calling it the black October lmao.
> node.js: security vulnerability.
> PHP: critical RCE vulnerabilities.
> tensorlake: npm package infected.
> liteLLM: malicious PyPI + Docker releases
> Railway: customer secrets exposed
> zeabur - infrastructure breached.
> OpenCode: AI API keys exposed.
> OVHcloud CLI: official binaries backdoored.
> RankMath PRO: malware inside official plugin.
> TeamUpdraft: customer data compromised.
> ConnectWise: cloud backup identities compromised.
> SubQuery: backdoored npm package.
> GhostAction: GitHub repos with malicious workflows (related to the broader Shai-Hulud/ChainDrop activity).
> eMudhra: certificate authority keys compromised (with downstream impact on Tailscale).
this some complete madness bro. 1/3 of the internet is fucked.
litellm 1.107.0 shipped to 23m weekly downloads with a credential stealer that runs at every python startup. you updated because not updating was the risk.
funny how every major lab shipped a proactive agent within days of each other. 15,570 views for the avalanche, not the product.
claude duet is a proactive agent that keeps working when you step away. 14,183 likes for a coworker who never rests and still has to ask permission first.
