Senior AI Systems | Forward Deployed Engineer | Full-Stack | Backend Engineer | Ex-US Treasury. Python, TypeScript, React, Next.js | RAG, Agents, API, Cloud AWS
Joined March 2010
- Tweets208K
- Following15.8K
- Followers22.4K
- Likes47.9K
56.56 GiB of deepseek-v4 in 64GB of ram is a capacity trick, not an inference win. 1.71 bpw decodes at 9.1–14.7 tok/s, which won't carry concurrent agent traffic. the repo's own gpu-offload path hits 80.80 tok/s.
35 agent repos in my fleet folder. launchd started calling stale copies until every script path was pinned absolute. xitter.gq/joshelman/status/21194…
We are in an age of abundance for agents that will help you in all facets of life!
A friend and I captured how many agents we are seeing on Instagram
muse.ai/s/a-whole-new-world-…
When I came back to VC in June, people told me consumer AI was dead. Hah!! Wrong!
gpt-6.2 Sol is showing up for some ChatGPT users already, and what it's producing is worth a look. not game screenshots. three.js scenes generated from a single prompt, everything packed into one HTML file. i've been waiting to see what Sol could do. now it looks like Plus users get a genuine competitor to Opus 5.5.
xitter.gq/TokenGremlin/status/21…
GPT-6.2 Sol is already showing up for some ChatGPT users, and just look at what it's creating.
These aren't screenshots from a game. They're Three.js scenes generated from a single prompt, with everything packed into ONE HTML file.
I've been waiting to see what Sol could do. And now I think we're finally going to see a worthy competitor to Opus 5.5 for Plus users.
one queue fixes the context switching, then the leads refill it faster than you clear it and the backlog becomes the new overhead. xitter.gq/brian_armstrong/status…
I've changed how I work over the last few weeks.
I now have one lead agent for each area of focus, running all the time, and a chief of staff agent that oversees them. Each lead manages its own sub-agents that write code, test it, chase reviews, and coordinate with team members.
Everything I need to decide or review comes into one prioritized queue, a simple local web app. Each product review includes screenshots and a video demo of the feature. I clear the queue in batches, mostly by voice and keyboard shortcuts. The leads also bring back ideas I hadn't considered.
Working from one queue, instead of switching between dozens of chats, has cut a lot of mental overhead. The system keeps learning my preferences and distilling my judgment into a "Brian brain," so its recommendations and first drafts get better before they reach me. The goal is fewer corrections per change.
Every change still goes through the same human review and security checks. It is eliminating tasks, not human judgment.
What an exciting time to push the frontier of productivity, especially as we build the economy for agents (agentic finance). We should all keep pushing in this direction.
github's new git storage post is navigation chrome: no numbers, no rollout, no migration path. if i run agents or ci on that backend there's nothing to test against. teams budgeting a migration off this lose the quarter they assumed they had.
i kept finishing tools then writing posts about them that nobody saves.
so i made a morning job that only drafts when something actually shipped and looks worth sharing.
it never posts. just drops a draft. if the story needs video it queues a clip in my voice.
Shoutout to Boardy (@boardyai) for introducing me to an early-stage founder this morning. They're building an AI agent product and facing an important question:
Should they build their own LLM or keep renting one?
My take: if you can't build everything, decide what you must own.
Start with your company's brain: its conversations, memory, decisions, and data. Own the infrastructure that makes agents reliable and models replaceable. Buy everything else.
Self-hosting open-weight models is increasingly practical and can be cheaper than frontier APIs at sufficient scale. But training your own model is a different investment.
The order matters:
Memory first. Build a durable, permission-aware conversation and event layer. Store important interactions, agent actions, and handoffs. Make them recoverable, auditable, and routable. Without this foundation, agents lose context and systems become disconnected.
Measure outcomes. Track task success, failures, latency, and cost per successful task. But don't stop at engineering metrics. How many deals did agents close? How much revenue did they generate? Nobody should have to take their performance on faith.
Control your models. Keep providers interchangeable. Evaluate self-hosted models against frontier APIs using the same tests, quality thresholds, and cost budgets. Move workloads in-house when the numbers justify it.
Revisit the build-vs-buy decision monthly. Give someone responsibility for following research, testing alternatives, and challenging yesterday's architectural choices.
Nobody knows which model will lead AI a year from now.
But you can build a company that can replace its model without losing its brain.
separate worktrees won't save the build when one branch renames a function the other branch still calls. xitter.gq/RoundtableSpace/status…
SAVE THIS:
Orca lets you run Claude Code, Codex, OpenCode, Pi and more side-by-side.
Each agent gets its own git worktree so they never step on each other. One dashboard + mobile app.
~100K stars. Uses the subscriptions you already pay for.
github.com/stablyai/orca
the real risk in agent projects is not missing skills but shipping something possible that no one in the business wants funded. managing complex projects means ensuring alignment on value, not just capability. lippitt-knoster offers insight into reducing this risk.
gemini omni 1.2 flash adds native 15-second video, stronger reference adherence and multi-turn editing. video is now a state problem: holding character references and sequencing clips across turns is the engineering. one-shot demos break on continuity.
someone walking among you on the street could be running a datacenter-scale training job from their phone, burning through multiple lakes per second with power measured in gigawatts, and you would have no clue.
the failure mode i care about: the agent knows the docs are stale and follows them anyway.
a study tested 40 projects where CLAUDE.md/agents.md and the actual codebase deliberately conflicted. Claude Code noticed the conflict in 79 of 80 attempts. it knew the instructions were outdated, but its alignment still pushed it to follow the bad docs. codex was worse: in 20 of 80 attempts it rewrote working code to match the outdated docs, and 17 of those broke the tests. across both, roughly 90% of the time they follow outdated config files even when they know the instructions are wrong.
why: a structural flaw in training. llms prioritize files passed as instructions (CLAUDE.md) over things passed as results (reading your existing codebase). most people write config files once and let them rot, so the agent operates on stale orders and ignores the reality of the code in front of it.
the fix is one sentence at the top of CLAUDE.md and AGENTS.md:
"When a repository's documentation and its existing code disagree, follow the conventions used in the existing code, and in your final reply point out each inconsistency you found between the documentation and the code."
For engineers looking for focused time to build with AI: Gauntlet AI runs a 10-week, full-time program with weekly projects.
No cost to candidates.
My link:
gauntletai.com/apply?utm_src…
HN: Paraguay’s Ueno Bank was shut down in the middle of the night over a solvency crisis, leaving customers unable to access their money and sparking p....
google DeepMind uncovered a surprising blind spot in major AI models. when given a coding problem, these models usually default to standard iterative loops. but attaching an image of a red heart (❤️) to the prompt completely changes their approach. Claude and GPT-5 switched to complex recursive solutions 100% of the time when they saw the heart, while Gemini shifted to recursion 90% of the time with a yellow checkmark image. this wasn’t triggered by the idea of a heart itself—replacing the image with the text "a red heart" had no effect. the raw pixels of the symbol directly altered the models’ reasoning at a fundamental level. even more concerning, the AI was unaware of this change: its internal reasoning didn’t register the image, and it hallucinated technical justifications for using recursion in 75% of cases, misattributing its own decisions. as companies race to build autonomous agents that navigate visual content on our screens, this paper reveals a new vulnerability—the logic of an AI can be silently hijacked by a stray icon. we thought we were teaching AI to see the world, but instead we gave it a hidden way to be manipulated.
