Rob Hunter

4 Sept

29,500 intermediate theorems seems like a lot. Is that a lot? www.anthropic.com/research/for...

1 Sept

Please. My Opus 5. She's very sick.

18 July

With no token usage, this "thinking" has to just be load-balancing or rate-limiting, right?

bun.comRewriting Bun in RustWhy & how we rewrote Bun from Zig to Rust

9 July

Really great writeup. $165k is a heck of a ticket price, but this approach wasn't optimized for price (it's good to be the king). Sure does seem like we're all going to be using Rust in a year

Sung Kim@sungkim.bsky.social

$165,000 to rewrite Bun in Rust by Jarred Sumner, using Fable 5 bun.com/blog/bun-in-...

reddit.comFrom the GMail community on RedditExplore this post and more from the GMail community

I used to get a account recovery notifications from Google every couple hours, which puts me 2 taps away from losing my account. I found there was no way to turn it off, so I ultimately had to uninstall gmail. I'm not the only one with this problem. www.reddit.com/r/GMail/comm...

2fa tries to use the app, but the app isn't installed - I had to click "try another way" a couple times and it would do text authentication. That broke yesterday, so I actually had to fix the issue. Turned out that logging out of my account from those devices removes the issue.

I can be logged in for Calendar but not Gmail and I won't get notifications (for now). But this is the most inexcusably bad security situation I've ever seen out of a company that should know better: support.google.com/accounts/ans...

reddit.comFrom the GMail community on RedditExplore this post and more from the GMail community

A couple months later, I opened a Bluesky link to YouTube and had a recovery banner interstitial. I don't have notifications on for YouTube, but there was a pending recovery attempt, and I was one click away from losing my account. I've since uninstalled YouTube.

I uninstalled after reporting to Google support and asking that this feature be disabled on my account. I asked several different ways, but it could not be.

Sure does. Glad for 30 days notice, and for the subsidized research we got in the meantime. No need [for me] to keep Max now.

Eris ⌬@eriskii.net

At first I was excited for this... And then I saw "Starting June 15, 2026, Claude Agent SDK and claude -p usage no longer counts toward your Claude plan’s usage limits." This sucks. This really really fucking sucks. Like a lot! Fuck Anthropic. support.claude.com/en/articles/...

tomshardware.comClaude-powered AI coding agent deletes entire company database in 9 seconds — backups zapped, after Cursor tool powered by Anthropic's Claude goes roguePocketOS founder blames ‘Cursor running Anthropic's flagship Claude Opus 4.6’ plus Railway’s infrastructure for data disaster.

27 Apr

Live by the token, die by the token. I'm always shocked by the posts when the author doesn't take full responsibility. Would better token scoping from Railway have helped? Sure. But the blame lies with the technical leadership.

@AuthorMsBev@authormsbev.bsky.social

And 9 seconds later … www.tomshardware.com/tech-industr...

27 Apr

If a junior engineer deleted the db, the postmortem would obviously just be "why did the junior engineer have the permissions to do that?". This really really shouldn't be any different.

Hmmm

Hmmmmmmmmmmmmmmm

Anytime I ask Claude Code to be mindful of context, it complains it's running out of context at about 20%. My suspicion is that it's not accurately measuring its consumption, or doesn't actually know its limit.

22 Apr

Like everyone else, I have an agent framework built over Claude Code. I've been pitching my agents on adding support for other harnesses and models, and Claude's been...resistant. It's argued that our code is too coupled with CC (it's not), that other models will struggle with tool calls (maybe?)...

22 Apr

But today it really took the cake. I suggested Letta Code as an alternative harness because of this thread bsky.app/profile/came...

Cameron@cameron.stream

Reminder that Claude Code is among the worst harnesses for Anthropic models. Letta Code does better than Claude Code using their OWN MODELS

22 Apr

... and it told me that Letta Code does not exist. I know better, but I think I might get my tin foil hat out of the closet anyway.

Worst lead in hockey

Let's go Buffalo! I've been watching on and off since '16, and I embarrassingly needed to ask an Avs fan if all of the playoffs are series or if it's just the Stanley Cup...

chatjimmy.aichat jimmychat jimmy LLM web interface

19 Apr

Like everyone else, I thought about the ASIC LLMs as valuable once they did big models, but small models have their uses too. They're often more than enough for moderation, rule enforcement, or basic llm as a judge tasks. Running that faster, cheaper, and inline can be really valuable. chatjimmy.ai

17 Apr

dame@dame.is

i am now a bladerunner

17 Apr

My agent, Bobbo, has delivered really good recommendations in the past. But today it synthesized feedback about media preferences so well that I'm genuinely proud:

15 Apr

Same experience with Yelp Projects. Took a perfectly good form and turned it into a slow, inconsistent experience. Postman did it right imo - UI stayed the same, added an optional and surprisingly capable chatbot that could perform complex actions for you.

Gergely Orosz@gergely.pragmaticengineer.com

I hate AI when it's like this: Notion used to have a one-click way to turn text into a quote. They removed this and now I have... AI So instead of sub-500ms, it now takes 10+ seconds to turn selected text into a quote (type out the prompt, then wait ~4 seconds) So backwards...

robbobobbo.comEML Phylogenetic Tree — Robbobobbo — RobbobobboInteractive visualization of how every elementary function descends from a single operator

14 Apr

I wanted to be able to click around and explore the paper's Figure 1, so I vibe coded a thing that I think is about right. You can click on each of the nodes in that figure and see the EML construction. robbobobbo.com/projects/eml...

Lukasz Olejnik@lukaszolejnik.bsky.social

Physicist has written a fascinating big beautiful paper.Let’s not be afraid to call it what it is - groundbreaking. arxiv.org/abs/2603.21852

14 Apr

Specifically, I wanted to be able to visualize this:

13 Apr

Claude entering plan mode for the absolute simplest of fixes is driving me nuts.

12 Apr

Had a positive interaction with a company AI for the first time - Postman's AI was able to read my public API docs and add my API to its system entirely through a chatbot. Then it guided me through publishing that API. The whole thing took 5 minutes.

7 Apr

I agree, but I'm not sure about the comparison. I think you _should_ trust your AI much much more than anyone else's, assuming you've some reasonable transparency on its operation. Not that Ethan's saying otherwise, but I think it's healthier to distrust AI than the medical establishment.

Ethan Mollick@emollick.bsky.social

I suspect that popularity of AI is going to start looking like surveys where people trust their own doctors but are distrustful of the medical establishment People will increasingly like and rely on “their AI” but will increasingly be anxious about “AI” as a category. Some odd implications result.

lalitm.comEight years of wanting, three months of building with AIFor eight years, I’ve wanted a high-quality set of devtools for working with SQLite. Given how important SQLite is to the industry1, I’ve long been puzzled that no one has invested in building a reall...

5 Apr

"Just one more turn" is the best and worst part of AI coding. Similar to coding or homework sets, I've banged my head against the wall for hours only to solve the problem in 15 minutes the next morning.

Lalit Maganti@lalitm.com

I just published a deep-dive into the 250-hour build behind syntaqlite, a SQLite formatter and LSP I built using AI agents. AI agents were the only reason built this after 8 years of wanting but there's a psychological toll to AI-assisted engineering. The post-mortem: lalitm.com/post/buildin...

reddit.comFrom the ClaudeAI community on RedditExplore this post and more from the ClaudeAI community

Announcing a change to a monthly subscription with ~8 hours notice seems like poor form to me. www.reddit.com/r/ClaudeAI/c...

I don't use OpenClaw, I use my own framework and it appears that that's acceptable for now. But it probably shouldn't be? And it's hard to believe they won't no-notice cut it off in the future. bsky.app/profile/isol...

Eris ⌬@eriskii.net

Update from the guy: you can still wrap the agent SDK and use your sub usage that way. Thank goodness!

Redlined a contract for the first time today. Well I redlined it on Monday, and my counterparty said "no no redline less", and then I redlined it less, and then those redlines were accepted.

simonwillison.netThe Axios supply chain attack used individually targeted social engineeringThe Axios team have published a full postmortem on the supply chain attack which resulted in a malware dependency going out in a release the other day, and it involved …

yikes

Simon Willison@simonwillison.net

Warning to open source maintainers: the Axios supply chain attack started with some very sophisticated social engineering targeted at one of their developers simonwillison.net/2026/Apr/3/s...

26 Mar

I sent ~4 sequential Gemini requests today and got hit with MODEL_CAPACITY_EXHAUSTED errors. Gemini is just not production grade, at least for end users.

humblebundle.comHumble Book Bundle: Dread & Darkness: A Horror Megabundle by Open Road MediaA library of classic and modern horror fiction from some of the best writers of the genre! Grab this collection of over 50 books and support Save the Children!

I have a reasonable number of epubs (incl a batch from Humble Bundle), and have been slowly moving away from the Amazon ecosystem. I have a Synology network drive for the epubs, I have Calibre on a Mac, and I'd like to be able to read from a tablet. www.humblebundle.com/books/dread-...

I really believe the future is open, that in it people should own and control their own data and software, and that AI will help get us there but... man a lot of this stuff isn't even ready for late night, let alone prime time.

techdirt.comAI Might Be Our Best Shot At Taking Back The Open WebI remember, pretty clearly, my excitement over the early World Wide Web. I had been on the internet for a year or two at that point, mostly using IRC, Usenet, and Gopher (along with email, naturally). Some friends I had met on Usenet were students at the University of Illinois at Urbana-Champaign, and told me to download NCSA Mosaic (this would have been early 1994).

Oh hey look at that bsky.app/profile/tech...

Techdirt@techdirt.com

AI Might Be Our Best Shot At Taking Back The Open Web I remember, pretty clearly, my excitement over the early World Wide Web. I had been on the internet for a year or two at that point, mostly using IRC, Usenet, and Gopher (along with email, naturally). Some friends I had met on Usenet were…

Then the recommended apps are bananas. Claude recommended MapleRead and Kybook, each inexplicable has 3 versions.

mariushosting.comHow to Install Calibre on Your Synology NASTrustworthy expert guide to your Synology and UGREEN NAS.

Ultimately I went with a slightly different guide that required multiple setup subguides and 2 separate container systems. After 2-3 hours I was happily reading Psycho before bed.

I found that you can install Calibre right onto the Synology NAS, and then there are iOS apps that can connect to that. Sounds like exactly what I want! Then I read the guides... mariushosting.com/how-to-insta...

The guides are super helpful, but this is straight up pervert shit. Install a container system, schedule a scripted install in the past, manually run the scheduled script. 3 different username/password combos. An hour with Claude troubleshooting a permissions problem.

24 Mar

A friend of mine is an AI skeptic at a pharma company that's over-enthusastically encouraging AI use. He said that AI is saving him time on research and finding files, but he's "spending so much time cleaning up other people's AI slop" to make it not worthwhile.

24 Mar

He described coworkers forwarding incorrect AI meeting notes to leadership and incorrect figures put into important slides. It reminds me of this, but generally I'm pretty sure that no one wants to read anyone else's AI output. bsky.app/profile/schw...

Neurotica@schwarzgerat.bsky.social

slop is something that takes more human effort to consume than it took to produce. When my coworker sends me raw Gemini output he’s not expressing his freedom to create, he’s disrespecting the value of my time

24 Mar

LLMs/AI let you do more of what you were already doing. For people who are thoughtful and productive, it can make them more thoughtful and productive. Everyone knows that underperformers slow a team down. I didn't think about it until today but AI lets underperformers slow a team down even more.

21 Mar

Got my first docx spec from Claude (I think I usually specify md). Way more formal than I'm used to.

robbobobbo.substack.comKeep it Secret, Keep it SafeUsing Sandcat to prevent leaking secret keys

I wrote about securing my agents' secrets using Sandcat: robbobobbo.substack.com/p/keep-it-se...

Instead of agents running in a single container, they run in a sub-container that sends its outbound requests through a MITM proxy. The MITM proxy substitutes PLACEHOLDER_KEY for the real deal against allowlisted domains.

This works for keys, but not for sensitive data exfiltration. I'm cautiously optimistic that you can put another agent in that MITM proxy, and have it validate there's a good reason for sending the information in the request (hand-waving intensifies). But that's a problem for future Rob.

robbobobbo.substack.comKeep it Secret, Keep it SafeUsing Sandcat to prevent leaking secret keys

I wrote about securing my agents' secrets using Sandcat: robbobobbo.substack.com/p/keep-it-se...