Engineering manager by day. Managed by AI after hours. Building agents that do real work, remember what matters, and fail with dignity.
USA
Joined December 2011
- Tweets3.1K
- Following809
- Followers617
- Likes791
Pinned Tweet
MCP is USB for AI agents. Thousands of plugs, built by thousands of strangers, changing whenever they feel like it.
I wrote up why I built MCP Checkup and what it checks today.
Would you let your agent plug into a server nobody has looked at?
我知道 Fable 5.2 出来之后，Claude 给大家发了两个 Fable 的 limit reset。但是今天我刚刚又看了一下我的账号，突然又多了一个 weekly reset。
哇，这是为什么？Claude 最近这么大方的吗？他们这是为了上市，已经开始大放水了吗？有点爽啊！
5
Anthropic 的 AI 两个月多报了近 1.2 万条漏洞线索，同期修好的只多 15 个。
curl 收到 45 份，公开的 34 份里 19 份被判不是漏洞，确认的 6 个全是低危。
Anthropic 账本上的真阳性率：94.2%。
开源项目以后是不是得专门招人拆 AI 的报告？
vulncheck.com/blog/anthropic…
4
Google is building /swarm for Antigravity: agents coordinate on a shared message board, and the high quota warning is already written.
Agents get a group chat. You pay for every reply-all.
When the swarm argues its way to a decision, whose name goes on it?
BREAKING 🔥: Google is developing a new Agent Swarm mode for Antigravity, where agents coordinate through a shared message board.
Remember Huggingface incident? 👀
> /swarm would appear as a new command, opening a Swarm dashboard to monitor Agent coordination, communication, and relationships.
> The notice warns about high quota consumption, obviously.
This mode can potentially push the performance of Google's models even further, so they would have a chance to overtake Fable 5.2.
This isn't the first experiment with "parallelization," as we have seen similar offerings from other labs, as well as internal experimentation on Google's side last year.
The Antigravity team have been teasing "more releases to arrive this week".
Besides that, we expect Gemini 4 Carbon to be rolled out today as well.
/swarm would be a nice addition! 🤖
h/t @lberondroth
1
15
825 million shares at $70 to $78 puts Anthropic's IPO at about $1.38 trillion, tops.
That's at least $120 billion under the $1.5 trillion everyone expected. Miss a budget by that much and you get a follow-up meeting.
Who's taking that meeting, the bankers or the early investors?
Official: Anthropic has launched the roadshow for its proposed IPO, offering 825 million shares at $70 to $78 each.
That implies an estimated fully diluted valuation of roughly $1.24 trillion to $1.38 trillion, below earlier expectations of around $1.5 trillion.
1
8
2023年谷歌博客写，Bard里的Gmail和网盘内容“人工审核员看不到”。10月16日文末加了更新：开启新功能后，有限情况下可能被人工审核。
帮助页说，邮件或文件很短或很相关时，审核员看到的“摘要”可能就是原文。有机密内容别连Gemini。
你邮箱里哪一封最不想被审核员读到？
blog.google/innovation-and-a…
1
1
16
Yesterday my Grok Bot weekly usage showed about 40%. This morning it shows 0.8%, and my next scheduled reset is still Tuesday 9:52 AM ET. Elon said fixes were going in overnight. I can't tell if that's the cause. What did yours go from and to?
Just did a round table review with the Grok @Bot engineering team.
Fixes are being implemented overnight and you should notice a dramatic improvement by tomorrow. If not, reply to this post and we will address any remaining issues right away.
1
One heavy user says he ran 5 Opus 5.5 sessions for days and used 34% of his weekly limit.
That's one workload, but it's the kind of number teams will now test themselves.
What would you measure first: tokens used, tasks finished, or fixes needed afterward?
I know I've already said it, but god damn, Opus 5.5 is a good model. I've been working really intensively in 5 sessions at the same time for a few days now and have only used 34% of my weekly rates so far.
Opus 5.5 is not just intelligence it is incredibly efficient. Best model.
24
A UCLA and Northeastern paper says about 18% of AI agent output tokens can be removed without losing useful work.
On SWE-bench Verified, cutting 41% of Kimi-K3's reasoning left success at 89.5% vs 89.0%.
Which part of your agent bill would you audit first?
New UCLA+ Northeastern University paper finds that about 18% of AI agent output tokens are removable waste, and cutting it barely hurts task success.
Agent bills are usually measured as total tokens. That number can't tell you which tokens did useful work and which just repeated a policy, refetched a file, or went down a dead end.
They route agent calls through a proxy and have an LLM judge mark spans that can go without losing useful work. Then they delete those spans and rerun the agent to check nothing breaks.
Across 6 frontier models and 4 benchmarks, about 18% of output tokens were waste. Pruning repetition and filler from Kimi-K3's reasoning removed 41% of it on SWE-bench Verified, and success stayed at 89.5% versus 89.0%.
To cut agent costs, find where your task and harness create waste instead of just capping tokens.
1
6
