julian (@herdingbots) profile picture

Head of Engineering Strategy. Building internal AI tools & DevEx by day. herding an army of AI agents by night.

Laguna Beach, CA
Joined May 2026
Photo by julian (@herdingbots)Photo by julian (@herdingbots)Photo by julian (@herdingbots)Photo by julian (@herdingbots)Photo by julian (@herdingbots)Photo by julian (@herdingbots)Photo by julian (@herdingbots)Photo by julian (@herdingbots)Photo by julian (@herdingbots)Photo by julian (@herdingbots)
So glad Jason said what we were all thinking about this scandal involving a super powerful male CEO cheating on his wife with three women, two of them his subordinates and one of them a probable Chinese asset, to devastating effect: this is because of DEI
RESET ALL BOARDS TO BE RANKED BY SHARES HELD. NO MORE DEI AND VIRTUE SIGNALING BOARDS NO MORE "EXPERTS" ONLY WARRIORS WITH SKIN IN THE GAME
42
48
2
1,058
41,656
Hi @Spotify, you traced my Chef’s Kiss illustration. It even has the wonky hand that I’ve regretted since I originally drew it. (There should be a pinky in the front instead of a thumb.) In a time when designers are getting less and less work, this really sucks to see!
164
3,821
71
65,777
1,243,143
julian retweeted
🌊💙 WE WERE ALWAYS BLUE
957
3,725
1,190
28,869
7,879,477
Compaction is a mistake. When an agent's context fills up, tools like Codex and Claude Code compact it: 1. The model prefills the entire session again 2. It writes a summary of everything so far 3. A new session starts from that summary I measured this on a real ~400k-token session with Opus 5.5: → 38 s and $1.99 when the prompt cache has expired → 29 s and $0.17 when it hasn't And the summary is lossy. The exact error message, the file path, the one number you needed is often gone. So I do something new for Mainplane Context retrieval works the other way around. Mainplane keeps the full record of every session as a file. When the context fills up, a new session starts right away with one instruction: "You are continuing session X. Retrieve key context from it as you go." The agent searches the old session and copies the exact lines it needs, when it needs them. Same session, same model: 15 s and $0.10 to get back to work. It does not depend on a cache hit, and nothing is paraphrased. This is the bitter lesson applied to agent memory. Don't hand-design what the model should remember. Give it the full record and good tools, and let it decide. Mainplane is open source (MIT). Link in the first comment.
Context compaction was a mistake The bitter lesson is to let the LLM manage its own context. Here is how I do this in Mainplane
312
154
123
3,594
562,336
julian retweeted
Native iPhone Duo apps, built with Expo. Ship yours before your AppleCare appointment!
368
803
419
11,746
1,114,309
julian retweeted
every product i use should just be a headless mcp, that feeds into an agent, where i never have to think about anything underneath. except for our product, of course. in our product, customers love the ui and love us very much. we're special that way.
150
412
109
7,093
339,223
This is retarded. Instead of reducing GDPR bloat regulatory shit they make small businesses jump through... They propose to let AI companies use your personal data to train AI. I am not anti-AI but this is duality of EU fake "citizen protection". dismantle the whole thing ffs (GDPR I mean).
‼️ EU governments approved a GDPR change to let companies train AI on Europeans' personal data without consent. It makes AI training a possible "legitimate interest", judged case by case. Parliament must still agree. Max Schrems calls it a "digital expropriation" of Europeans.
40
44
1
474
31,817
Here is official trailer, You Can See Everything
2,200
13,147
5,737
179,350
39,485,091
Did a write-up of my own German exit tax solution. TLDR: Needed 13 tax advisors, 1.5 years, and 18.7k€. Tax advisors shady as hell. Highlights: You have to pay 2k€ to a notary to move your companies into a new structure. The notary is missing the right document template. You have to pay another tax advisor 1.9k€ for said document template. After paying 1.9k€ for the template, the notary reads it out loud to you and you pay them another 2k€ for that
44
18
5
460
72,613
julian retweeted
efficient markets in shambles
JUST IN: NVIDIA's $12,930,300,000 acquisition of Hugging Face contains an easter egg. The number 129,303 is the decimal conversion of Unicode point U+1F917. The 🤗 emoji.
87
262
12
7,649
551,369
A website for trading fake money is a net negative.
Coinbase CEO Brian Armstrong says donating to charity is net negative "The common view is that philanthropy is this noble cause, and Bill Gates is like 'I'm going to go rehabilitate my image and do philanthropy.'" "A lot of charities and philanthropies are actually net negative for the world and they get captured. It's remarkably hard to find a foundation that has not gotten captured by ideology.” “By trying to do something good, like criminal justice reform, it turns out they just increased crime. I have a donor advised fund and I put money toward things helping civilization advance, but I don't think at this moment I would ever create a foundation." "If you look at the Ford Foundation, I think Henry Ford would be turning over in his grave if he knew what they were up to.” “I have a DAF, a donor advised fund, and I just try to put money toward things that I think are helping civilization advance. But I don't think at this moment I would ever create a foundation."
45
116
7
2,182
119,198
I like using murder rates from medieval Europe to get at how insanely violent places American cities are. Baltimore has a similar homicide rate to... Italy, during the time of the Guelphs and Ghibellines, when Italy was a bunch of intermittently warring states.
Chicago's homicide rate is completely absurd.
200
689
65
6,449
516,157
"I don't want invisible watermarks embedded in my content" then........ write it?? Yourself????
🚨 JUST IN: Claude models will now have invisible watermarks embedded in ALL text, and ALL metadata attached to files…
39
1,622
8
15,922
205,852
claude scraped the whole internet of content without attribution or compensation for its model but wants credit for its work
Anthropic says new Claude models will embed invisible watermarks in all generated text, everywhere Claude is offered. The watermark is part of the text, it isn't metadata: "it will travel with the text when it's copied and pasted elsewhere, and may persist through some editing." This starts with models launched on or after August 2, 2026, under an EU AI Act code Anthropic signed. Anthropic is still working on adding it to current models. The rollout is worldwide.
171
4,423
90
46,369
1,014,355
julian retweeted
I don't always clear my calculator... but when I do, I hit both 'C' and 'CE' a bunch of times because I don't know exactly what they do.
286
1,420
70
60,532
2,041,360
julian retweeted
I’m sorry but a model escaping a “sandbox” says more about the sandbox than the model
256
436
85
8,480
224,774
If you're ever giving a talk on how difficult it is to do data anonymization well, this is your lead slide.
65
129
28
3,783
210,170
Big Four accounting firm Ernst & Young has just been leaked by ShinyHunters
39
173
32
3,059
883,286
VCs have scrutinized the unit economics of Mamdani’s grocery stores more closely than those of any startup founded in the past four years
15
59
6
1,098
44,287
Funny to see the same people defending Leopold for "being in the arena" and losing his investors' money criticizing Airtable and Howie for actually being in the arena for 14 years and returning their investors' money (and getting some for employees).
4
6
66
3,660