runesleo

Meta’s engineering team shared a counterintuitive conclusion: the more knowledge an agent has, the more likely it is to make mistakes.
They initially built an omniscient agent that knew everything. It performed normally in demos, but went off the rails in production: applying System A’s naming conventions to System B. The solution wasn’t to write better prompts, but to make the agent handling orchestration know nothing: 1,300 lines of code, responsible only for classification, collecting fields, checking, and assigning tasks, while 14 specialized agents each handled one area and could not see

ChatGPT quietly changed its frontend again yesterday. I have a script that automatically sends commands to the web version, and it had always relied on the #prompt-textarea ID to locate the input box. Yesterday, it suddenly could no longer find it—they replaced the input box with a ProseMirror rich-text editor, and that old ID was gone. Changes like this come with no announcement; you only find out when your automation breaks. The lesson is: for any automation that relies on a webpage DOM, don't hard-code a single ID. It's best to keep a fallback selector or two; otherwise, the moment the othe

Yesterday, an entire batch of my automation scripts suddenly stopped working. After investigating for ages, I found it wasn't a code issue—the brew upgrade had automatically relinked python3 to 3.15, and the new version's site-packages was empty, so none of the dependencies previously installed with pip were there anymore. Even worse, the scripts didn't report any clear missing-dependency errors; imports just failed for no apparent reason. I later added a check to the monitoring: verify once a day that python3's dependencies are still present, so any brew change can be detected immediately. Re

Open Question | TapeOut Developer Acceleration Program: What Tools Would People Actually Use?
The X Layer × TapeOut Developer Acceleration Program is open (deadline: UTC 12/7). Submissions must use TapeOut’s circuits, contracts, on-chain data, or DeWEB, and clearly explain who would pay for them.
I’ve been playing around with Fab X, and the most tedious parts are monitoring output, selling in batches based on price, and adjusting circuits. I’ve written scripts to automate all of these. So I wanted to ask everyone:
· Which step is the most annoying when you play, and the one you’d most like som

There’s been a claim going around: X changed its ranking system in October, and posts containing token symbols like $BTC are treated as bots, with impressions dropping by as much as 80%.
I compared 370 of my own posts from early September to mid-October:
· With $: September median impressions were 2,085, versus 2,470 in October
· Without it: 1,849 in September, versus 1,447 in October
At least for me, posts with token symbols were not suppressed—in fact, they performed better (there were only 10 posts with $ in October, so the sample is small).
What actually declined was the overall figure, a

After using Claude Code for a while, three things are most annoying: the title stays stuck on the first sentence, it’s unclear which conversations can be closed, and no one flags matters that only I can decide.
I wrote three hooks: remind me to retitle when the main thread changes; write one line of “archive check” at every wrap-up; put the “needs my decision” list in front of the model, forcing it into four categories with the totals adding up.
It uses only the Python standard library, requires no internet connection, and works after copying two files and tweaking the settings.

@ellipsefun's bridge custody on Arc is empty.
Their CRCL vault on Robinhood Chain (0xed811A67C40B25D02A39492D91Aaab335b2cBD92) holds 0.08 CRCL against 241 bCRCL issued. Their own reserves page still shows "100% coverage".
Where it went:
- Oct 10 & 13: the vault sent 556 CRCL to 0xfef5e10566b0829a9e266f2973a74678ebfe6d6d, the same address that collects Ellipse's bridge fees.
- That address sold 285 CRCL for 24,022 USDG (0xd5373c6b87da44f3564224d841186f5d3499c699d551bb30c978a0b556f8a18d) and 72 CRCL for 6,164 USDG (0x19bc0ebf8c64190568335833b0916c03e4534c288517c2fe1cb607b4bf1aada6).
- Oct 9: the

Letting AI help me get work done with ChatGPT has a side effect: a new conversation is opened for every task, and the sidebar is packed full by the end of the day.
Added one step: after saving the answer, AI clicks “More → Archive” to put the conversation away. It doesn’t rely on coordinate-based clicking, but directly finds the button on the page; if rate-limited, it skips without retrying; and even if archiving fails, the task result is unaffected.
If AI is going to do the work, it should also clean up after itself.

After using Claude Code for a while, three things get annoying: the title stays stuck on the first sentence, you can’t tell which conversations can be closed, and no one flags the decisions only I can make.
I wrote three hooks: remind me to retitle when the main thread changes; write one line—“Archive check”—at the end of every session; put the “Needs my decision” list in front of the model, forcing it into four categories with matching totals.
It uses only the Python standard library—just copy two files and tweak the settings to use it.

Although I’ve tried various AI products, such as ChatGPT, Claude, Grokbot, and Cursor, I still prefer having just one conversational interface.
I want to get all my work done through a single conversational interface, without having to switch back and forth. I wonder if anyone else feels the same way? For example, setting up conversations with various bots in Bot may feel great at first, but over time, it can really become exhausting
So what I’ve always been pursuing is opening a new ChatGPT conversation or a new Claude conversation, and then having it gradually help me solve all my subsequent

The most annoying thing about using Claude Code, Codex, Cursor, and Grok at the same time is that something discussed with one can’t be found when switching to another.
I’ve open-sourced the export script I use: it exports local chat histories from all four into Markdown, one file per conversation, so any AI can find them with grep or full-text search.
· Reads local files only; no network access and no changes to any app
· Incremental exports, processing only conversations that have changed
· Replaces credentials such as API keys and tokens before writing to disk, so the files can only be read

Yes, GPT-5.6 Sol can now be selected directly in Cursor. However, OpenAI plans to discontinue direct access in Cursor around November 12, after which you’ll need to enter your own API key to continue using it.

When you ask ChatGPT or Claude to help execute on-chain operations involving funds, they often refuse, and nothing you say will change their mind.
At that point, you can try Cursor App: have ChatGPT or Claude write the prompt for you directly, then copy it over and execute it. Auto works—tested personally