More to throw in the Big Deal pile openai.com/index/openai...
Who's Who of AI
Phillip Carter
morbjs maintainer
work: salesforce, past @honeycomb.io and @microsoft.com
play: computers, backpacking, snowboarding, hiking, kayaking
What they're sharing
openai.com
Evaluating AGENTS.md: Are Repository-Level Context Files Helpful for Coding Agents? arxiv.org
CrankGPT — fully offline, human-powered local AI squeezlabs.github.io
Do people click on links in Google AI summaries? | Pew Research Center pewresearch.org
AI in Search is driving more queries and higher quality clicks blog.google
From Storage to Experience: A Survey on the Evolution of LLM Agent Memory Mechanisms arxiv.org
openai.com
interconnects.ai
The System Is the Asset - The Phoenix Architecture aicoding.leaflet.pub
I don't know how to evaluate coding models phillipcarter.dev
Agentics / Tech Things: Tokenmaxxing is dead, long live tokenmaxxing 12gramsofcarbon.com
I trained a transformer in an email - Adamos Hadjivasiliou adamoshadjivasiliou.com
Articles & links
Well this reads as quite the combo breaker, at least with agents powered by LLMs ~3 months ago, developer-submitted and LLM-created AGENT md files don't seem to improve task performance for python projects, but do burn ~20% more tokens arxiv.org/abs/2602.11988
AI Weekly's analysis
→
- Context files like AGENTS.md tend to reduce coding agent task success while raising inference cost by over 20% on average, the paper reports.
- On a new 138-issue Python benchmark, LLM-generated context files added 3.92 steps per task and pushed costs up by 23%.
- The finding held across Claude Sonnet-4.5, GPT-5.2, GPT-5.1 mini and Qwen3-30b-coder, so it is not tied to one model family.
Read full analysis →
Holy shit this is awesome, hand-cranked LLM inference squeezlabs.github.io/handcrank/
AI Weekly's analysis
→
- CrankGPT runs a full voice-interactive AI pipeline on a Raspberry Pi 5 with 8GB RAM, powered solely by a 20W hand-crank generator.
- Cold-start to functional conversation takes roughly 30 seconds; time to first token ranges from 0.8 to 2.9 seconds depending on model size.
- Memory bandwidth, not raw compute, is the primary bottleneck for on-device LLM inference, with DDR5 hardware achieving 29-58% faster token generation than DDR4.
Read full analysis →
Some early research definitely shows less click-throughs, but it doesn't say much about if the clicked link was a garbage site or not. In theory with wikipedia, .gov, reddit, etc. showing up more in sources there's more opportunities for better content I guess www.pewresearch.…
AI Weekly's analysis
→
- Pew found Google users click a traditional search result in 8% of visits when an AI summary appears, versus 15% without one.
- Users clicked on a link inside the AI summary itself in just 1% of visits, per Pew's 68,879-search dataset.
- Browsing sessions ended on 26% of pages with an AI summary, compared with 16% of pages without one.
Read full analysis →
I'm curious how this meshes with their own (generic) words, where they claim they're sending clicks to ostensibly better content than before. I don't know if I quite buy it, but there's also so many unreliable narrators around that it's hard to know for sure blog.google/produc…
Survey papers are so funny sometimes. Like this one about memory systems for LLMs that has to hamfist in some math expressions at the start and then never use them arxiv.org/abs/2605.06716
Huge news for openai openai.com/index/openai...
I’m personally not holding my breath on the prose improving, sadly. More being forced to read sloppy PRDs for work, until morale improves I guess www.interconnects.ai/p/why-ai-wri...
A related good post by Chad Fowler aicoding.leaflet.pub/3mbp5ukeuzs22
Showing my whole ass to you all and admitting that I can't keep up with nor evaluate coding models on my own www.phillipcarter.dev/posts/i-dont...
This was a good read on the org dynamics side of things: > It’s now a few months later, and the tokenmaxxing policies had their intended outcome: everyone is using AI to code, at least a little bit. 12gramsofcarbon.com/p/agentics-t...
I disagree with the author, the future of LLMs _is_ using emails as compute www.adamoshadjivasiliou.com/blog/transfo...
Their own posts
Recent commentary
Oh not a whole lot of PMs in my follows but two bad things are currently surging in PM world: 1. Writing overly verbose specs/PRDs with AI, not even bothering to get creative with prompts 2. Using claude code to spit out a hifi gestalt of a UI prototype and dev teams getting pressured to ship
Still stunlocked that Marc Andreessen and his big brain prompts AI models with “don’t hallucinate”
My dumb AI take of the day is that I think it was wrong to introduce 1M+ context windows on account of coherence != capacity
What's the state of the art on coding agent memory these days? I haven't used it much, at least not with Claude. So far the best memory implementation I've seen is just normal ChatGPT. It seems to do retrieval quite well.
I hadn't realized how easy it is to get started with base llama.cpp these days. I gave up on ollama a while ago since they started getting weird. Anyways: brew install llama brew install pi pi install git:github.com/huggingface/pi-llama llama serve <model> pi Got muh gemma4 on muh macbook
I think there’s something to it that Ed Zitron gets a lot of readers who love his overly verbose writing … and that a lot of people get impressed with the overly verbose words barfed out by LLMs in default settings These groups are nominally non-intersecting but I don’t buy it
Yeah so like, I know it’s all hip to doom-jerk about how nobody is going to hire juniors or middle managers anymore because of AI (directly or as an excuse), and I think it’s high time to invest in the counterfactual here, because it’s gonna hit faster than you think
Does giving an LLM a persona in the system prompt do anything these days? Like the "you are an expert coder" in pi's system prompt, does that do anything? What about for other kinds of tasks?
So anyways I read Amodei’s post and he’s put forward what I think is the most complete policy framework for AI that doesn’t hurt OSS developers either. That the current US Gov of dipshits decided to yoink Anthropic’s latest model just feels like retaliation for him not capitulating to the DoW
June is gonna be a month with a lot of bad math takes because of Copilot moving towards usage-based pricing
Their network
In Phillip Carter's orbit
Center = Phillip Carter. Left = members they follow (green edges). Right = members who follow them (blue edges). Top = mutual follows (orange edges, slightly larger). Drag any node to reposition; click to open that profile.