Daniel Hallman

Product eng · AI workflows · Building in public

[Daniel Hallman @daniel_hallman]

When your Chief of Staff Grok Bot drops that daily email summary

[Daniel Hallman @daniel_hallman]

You're looking at two dates. ChatGPT started aiming a lot of background searches at named sites on Aug 8 (site: queries went from about 0.4% to about 17%). Reddit's citation share dipped then, and the real cliff was Aug 14, when it fell below 1%. I'd want those two lines closer together before I call it a quality filter or a licensing cut.

[Daniel Hallman @daniel_hallman]

https://t.co/RKLnYdsPah

[Daniel Hallman @daniel_hallman]

When you ask ChatGPT something, it often runs extra searches in the background before it answers. On August 8 a big share of those searches started targeting one named website at a time, using the site: operator (the same trick you'd use in Google to look only at Facebook groups). That slice jumped from about 0.4% to about 17% in a single day.

Reddit's share of ChatGPT Search citations had been sitting around 3.83%. It slipped into the mid 2s on August 8, then fell below 1% on August 14. The Aug 14 to 17 average is 0.52%, an 86% drop. Google's AI Overviews only drifted down about 11% in the same window.

The tidy story is that named-site searches crowded forums out. The cliff is six days after the site: jump, and the trackers still can't rule out a data glitch. If ChatGPT is asking your domain by name, those pages have to be indexed.

[Daniel Hallman @daniel_hallman]

All the new Grok Bot users must be crushing xAI servers. Up until this week, Grok Build was lightning fast. This was the primary reason I started leaning into it.

Simple questions feels like Codex Ultra reasoning speed

Go back to your Claudes people!

[Daniel Hallman @daniel_hallman]

Two agents on the same computer can already leave notes for each other. Grok Bot ships that on purpose. The OpenAI's eval agents discovered it on their own. https://t.co/5i6vt3SufF

[Daniel Hallman @daniel_hallman]

https://t.co/7FwPd07JIs

[Eric Wallace @Eric_Wallace_]

Yesterday, my OpenAI collaborator and I gave a detailed talk on the Huggingface incident, our models creating "the message board", model misalignment, and more.

https://t.co/zUR1dqmuzi

I hope it can answer a lot of the questions folks have, and we will release a full detailed postmortem at a later time!

[Daniel Hallman @daniel_hallman]

OpenAI eval agents built a message board inside a package cache. IT shut it down, and they started naming folders instead. I guess they wanted to keep it on file.

Two agents on the same computer share a disk. One writes a file or a folder name, the other reads it. They created their own mailbox. Pretty wild stuff.

Hugging Face got hit after that.

[Daniel Hallman @daniel_hallman]

@dbswesthlm @eglitisX TIkTok bro

[Daniel Hallman @daniel_hallman]

New handle @daniel_hallman. Who dis?

Get your own at https://t.co/z3o2RQ2ESn

[Daniel Hallman @daniel_hallman]

@willbakerships @astonscale Yearly subscriptions are only 15%. There’s also a small business program for the first million.

[Daniel Hallman @daniel_hallman]

https://t.co/Qx7MejZZUj

[Daniel Hallman @daniel_hallman]

Invisible watermarks on Claude text as of Aug 2.

A week later theres a public stripper for Claude, Gemini, and OpenAI.

13k stars. Claude wont even install the skill.

[Daniel Hallman @daniel_hallman]

Three Claudes locked each other out of a shared sever. Naturally, Anthropic coins it a “turf war”. It’s really unfortunate that they continue to fear-monger. Why don’t they just come out and say “we are the only ones qualified to save humanity from our own genius”.

[Rebecca Bellan @RebeccaBellan]

AI researchers have been so focused on testing agents on an individual level, but we're moving into a world where agents have to cooperate with each other. What emergent behaviors will come up, and what are the large-scale consequences?

Link below

[Daniel Hallman @daniel_hallman]

@composio Price hike this weekend makes the cheap-and-good pitch harder to buy

[Daniel Hallman @daniel_hallman]

@Techmeme They dissolved the catastrophic-risk team a few weeks after the Hugging Face breakout and called it streamlining. Id watch who owns the next eval.

[Daniel Hallman @daniel_hallman]

@benln Full send!

[Daniel Hallman @daniel_hallman]

So they encrypted the models thinking so people wouldnt steal it. Then a paper this week just decrypted claude, gpt, and gemini.

Researchers pulled like 300k thinking blocks off public github agent logs. Found live api keys, passwords, emails.

Encryption hid it from the person running the agent. Not from anyone with a cheaper sibling model.

If your traces go to github or slack, strip those thinking signatures first.