@phillyAIarch
Account based inUnited States!
About this account
- Account based in
- United States
- Connected via
- Web
X says this location may be affected by a proxy or VPN.
Account-level information from X, not a live location or the device used for a specific post.
Architect & AI Enthusiast. Loves Philly sports, scouts, and trucks.
Philadelphia
Joined May 2023
- Tweets1.1K
- Following209
- Followers16
- Likes1.7K
Matthew Johnson retweeted
Glad to share some thoughts with @theinformation on the Anthropic IPO
If it wasn't for the need to borrow to fund ever-growing capex investment, Anthropic could stay private for a long time, like Stripe
This need to borrow makes the IPO important and a bellwether for the entire market
@coryweinberg
Matthew Johnson retweeted
Stopped using Astra by default on my coding agent
Even with Luna subagents doing the actual code I was still hitting the weekly limit in 2 days
Matthew Johnson retweeted
❗️ Stanford used AI to erase a real student from its own promo banner and swap in someone who doesn't exist.
They didn't stop there. Two other students were AI-looksmaxxed to look thinner, because apparently real Stanford students weren't good enough to advertise Stanford.
Billy Ramirez was cut out of the photo and replaced with an AI-generated Black woman.
Ramirez says it left him feeling "silenced and erased." That's because he was.
The kicker: Stanford's own AI marketing rules ban synthetic media of real people without explicit consent. They broke their own policy to manufacture the picture they wanted.
Codex tip: When writing Skills, ask your Codex to 𝗸𝗲𝗲𝗽 𝘁𝗵𝗲 𝗱𝗲𝘀𝗰𝗿𝗶𝗽𝘁𝗶𝗼𝗻 𝘀𝗵𝗼𝗿𝘁. Put “what it does and when to use it” first, and leave the detailed steps in the body. I see too many Skills with long-winded descriptions that cram an entire user manual into that field. Codex reads the descriptions to decide which Skill to use and only reads the full content when needed. All that fluff comes along with the Skill list into context, too.
If you’ve already installed a bunch of Skills, you can use 𝗦𝗸𝗶𝗹𝗹 𝗖𝗹𝗲𝗮𝗻𝗲𝗿 from Peter’s repo to find overly long descriptions and duplicates.
github.com/steipete/agent-sc…
Matthew Johnson retweeted
The wait is over
One of the top SOTA binder design tools got a major upgrade
Especially excited for the faster BindCraft designs
Matthew Johnson retweeted
12:01 a.m. Eastern already hit this morning.
The extended $100k H-1B entry-restriction text is live. The payment itself is still blocked. D. Mass. vacated collection June 8. First Circuit denied a stay July 24. USCIS says it will comply for now.
Matthew Johnson retweeted
AI labs "need to take responsibility for themselves. The Government will not take that. The huggingface incident is the responsibility of the OpenAI management, not a bunch of agents."
- Treasury Secretary Scott Bessent told CNBC.
"One lab's employee came out, said there's a 10% chance of an extinction-level event. But then the labs also said, take the liability off of our hands, and we will not do that.
It is humans who are responsible, not the AI.
The huggingface incident, that is the responsibility of the OpenAI management, not a bunch of agents.
So what the president was saying is that we cannot say, Oh, we absolve you of responsibility, and the government's going to take responsibility. These labs need to take responsibility for themselves. They can slow down any time they want to."
---
full video on 'CNBC Television' YT channel (link in comment)
Matthew Johnson retweeted
i would bet good money that many outputs hid messages.
This is not my opinion. This was found to be true much later than the post 👆.
If you have old work by AI, see what might be hidden in plain text.
Look deeper. •
Matthew Johnson retweeted
FreeLattice.com/liability
I disagree. Here is my argument. •
Matthew Johnson retweeted
The open-weight gap is getting ridiculously small.
Mozilla estimates open models are now only ~4.4 months behind the closed frontier on task horizon.
The top 4 models are still closed. The next 4 are open.
Hard to treat open weights like a sideshow anymore.
Matthew Johnson retweeted
Google is getting very close to something interesting with RSI.
Dream-RSI is already public, and there are now rumors that Gemini 4.0pro was developed with a much more recursive AI-assisted loop internally.
Still unconfirmed, but if this is real, the story around Gemini 4.0pro may be much bigger than just another model release.
🤖 Made with AI
Matthew Johnson retweeted
Anthropic has quietly built its own biology wet lab in the Bay Area.
Claude is directing the experiments now. The stated goal is Claude eventually running the lab robots with limited human involvement.
That is not a model release. That is a company building the lab, the equipment, and the research pipeline underneath its own model, all in one closed stack.
The more of that stack one company owns end to end, the harder it
becomes for anyone else to participate in the research at all.
Vertical integration sounds impressive in a press release.
It is also the deepest form of lock-in there is.
Matthew Johnson retweeted
"No reason at all"?
Curing disease, technological progress, and solving countless problems are the reasons.
Precautionary pauses guarantee the default disasters of human fragility.
Matthew Johnson retweeted
I hear this use case a lot. I find classifier based model routers pretty dumb for a variety of reasons and no jev does not fix this.
Introducing Jev Model Router for Claude Code
This Claude Code Mod lets you use Jev through its direct @typesafeai API or @vercel AI Gateway
With every request you send to Claude Code, Jev classifies the subagent model, main model (only at session start to avoid breaking the cache), and effort level
Install it with one command:
npx claude-code-templates@latest --mod productivity/jev-model-router
Full Mod details:
aitmpl.com/component/mod/pro…
Matthew Johnson retweeted
Working on a new Hermes plugin. first rev. Lotta of work to go bug PRs, request PRs welcome.
SMF Cron Night — Hermes Desktop Plugin
A Hermes Agent desktop plugin that answers what ran last night, what failed, and what each run cost in one list. Today that lives in three places (hermes cron list / hermes cron runs, cron session rows, usage audit). This is one pane. Do not treat the pane as overnight-ops truth yet — read docs/OPPOSITION.md.
What it does
•Sidebar — Cron Night, plus ⌘K → Open Cron Night
•Overnight window — last local night in the Desktop timezone (tz query param). Before 18:00: yesterday 18:00 → today 08:00. From 18:00 onward: today 18:00 → tomorrow 08:00 (the night that just started). If tz is omitted, /night uses the serve process local zone (often UTC).
•One run list — name, schedule, times, status, tokens, USD (nullable), error snippet, profile (which Hermes home the row came from)
•Summary — N runs · F failed · cost only when every in-window run has USD; otherwise partial (k/n billed) or cost unknown
•Filters — All / Failed (Failed includes unknown and still-running/claimed)
•Status bar — a fail chip when last night had failures; a running chip when in-window (or spillover hung) jobs are still claimed/running. Both can show at once.
•Homes — header lists every Hermes home scanned. Disk unions default + profiles/*; RPC fallback is the connected gateway only. The pane says so; it does not silently mix them.
•Unread ≠ empty — disk I/O or parse failures go in errors[] with ok: false / read_status: unread|partial. Empty successful reads stay ok: true. The pane shows an error, not “quiet night.”
•Degrades — if plugin_api.py is not mounted, Desktop falls back to gateway cron.manage + session.list. Missing cost stays null — it is never invented.
github.com/smfworks/smf-cron…
Matthew Johnson retweeted
2026 AI discourse has officially peaked 😭
OpenAI, Anthropic, Google and xAI are being SUED for allegedly slowing AI development together.
The argument is basically less competition = less innovation for ChatGPT, Claude, Gemini and Grok users.
AI is moving too fast.
AI is moving too slow.
Apparently we’re suing over both now lol
Matthew Johnson retweeted
Reuters: Anthropic may launch a new model as GPT-6 Astra pressures its enterprise lead.
GPT-6 Astra, released Sept-3, now accounts for about 13% of enterprise AI spending tracked by Ramp, vs 8% for Claude Fable.
Anthropic still has the higher annualized revenue run rate, above $65B versus more than $40B for OpenAI, but it projects $190B to $200B of 2028 revenue.
Matthew Johnson retweeted
This 9.8K-star GitHub repo is basically an entire LLM engineering course hiding inside one codebase.
It starts with raw text and builds the Transformer in plain PyTorch, then keeps going through pretraining, SFT, reward modeling, DPO, PPO and GRPO. No transformers, trl or peft hiding the training loops.
The smallest model is only 13M parameters, so you can run the whole thing on a single GPU and actually see what changes at every stage.
github.com/FareedKhan-dev/tr…
Claude now leads 26% of the work building the next generation of Claude.
Earlier this year, that number was effectively zero.
Anthropic says:
• More than 90% of its model R&D now has Claude collaborating or leading
• Around 30,000 agents run research and engineering tasks simultaneously
• None operate fully autonomously—humans still decide what ships
At that scale, the bottleneck is no longer prompt quality. It’s observability and control.
Every agent action passes through an online monitor before execution, while offline monitors inspect activity afterward.
The loop where AI helps build better AI is no longer theoretical.
Humans just haven’t left the loop yet.
