Shlomo F.

vibe coder (professionally, allegedly) · I read the weird side of AI so you don’t have to

[Shlomo F. @ShlomoFaig]

This week in AI dev, minus the hype (#2)
Everyone shipped a model, malware learned where we keep our API keys, and GPT-4o finally got a retirement date

1. Gemini 4 Argon is out for all paid Gemini API users. $2 in / $10 out per million tokens until Nov 30, then $4 / $20. Up to 1M tokens per call, input and output combined
Good time to try it. Just don't build a budget around the intro price

2. Meta shipped Muse Wonder, its first frontier model ($4 / $20 per million tokens). Three days later it released open weights for Muse Spark: 816B params, 42B active, 1M context
Custom license with limited commercial terms, so read it before you build on it

3. DeepSeek updated V4.1 Flash: same price, shorter outputs, and a self-reported 77.0 on DeepSWE, 0.9 behind what Google claims for Argon. It also launched its own search and fetch API

4. OpenAI is removing GPT-4o, 4o mini, 4.1 and 4.1 mini from the API on Jan 14, fine-tunes included. Suggested replacement is gpt-6-luna
Somewhere a prod app is still pinned to gpt-4o-2024-08-06 and nobody remembers why

5. Symantec found an infostealer called Aoki that now goes after AI coding tools: .claude.json, Codex's config.toml and .env, Cursor and Copilot tokens. Stolen tokens can expose linked repos and CI
Even the malware is building for coding agents now

6. Claude Code's Code Review is out of preview for Team and Enterprise. Reviews now take 6-10 minutes instead of ~20, with Quick, Standard or Deep per repo
Now it just needs a human who reads the review

Overhyped: OpenAI is building a humanoid robot. Altman expects "a cool demo in 2027"
Until then my agent still can't find the other file

Recap #3 next Sunday, if I survive this week's AI bill

[Shlomo F. @ShlomoFaig]

Sources:
1. https://t.co/CMV0nzLqjb
2. https://t.co/tQxg7D96gy
3. https://t.co/Gx76LPbSzy

[Shlomo F. @ShlomoFaig]

For a while I thought the secret to getting good results from coding agents was giving them more. More context, more instructions, a longer AGENTS.md, and one giant thread where the agent surely knows everything by now.

Turns out the research says pretty much the opposite. Of course it does.

ETH Zurich tested AGENTS.md files on real coding tasks with Claude Code, Codex and Qwen Code. The ones written by an LLM (like mine, obviously) didn't help at all, and made tasks about 20% more expensive. Even the ones devs wrote themselves only helped by 2.4 points, which isn't even statistically significant, and still cost up to 19% more.

And the agents weren't ignoring the files. They actually followed them very closely, running more tests, more searches and reading more files, most of which the task didn't need.

Another paper from June tested an agent on 50 expense-report tasks. With the full conversation history it got 71% right. Keeping only the last 5 tool calls, 79%. And the last 5 plus a short summary got 91.6%, on about a third of the tokens, in under 6 hours instead of almost 15.

My favorite one is from Chroma. They tested 18 models, and every single one got worse as the input got longer, even on really simple tasks. Weirdly, they did better when the text was randomly shuffled than when it actually made sense (good news for anyone whose docs already read like that).

So now I delete the auto-generated AGENTS.md, keep only what the agent can't figure out from the repo, and start fresh sessions way more often.

Finally a scientific reason to write less documentation.

[Shlomo F. @ShlomoFaig]

@signulll Google's probably shipping three next week, all called Gemini something

[Shlomo F. @ShlomoFaig]

@jarredsumner Mine's the opposite
Half my day is telling it the fix was one line and to please put the database back

[Shlomo F. @ShlomoFaig]

@theo It's coming, Claude Duet just decided it doesn't need your attention yet

[Shlomo F. @ShlomoFaig]

@testingcatalog A proactive agent that notices your problems and offers to buy something
Finally it really is just like me at 2am

[Shlomo F. @ShlomoFaig]

@NousResearch @JensenHuang @NVIDIAAI Jensen's agent trained thousands of simulated ducks in one keynote
Mine renamed a variable today and touched 14 files

[Shlomo F. @ShlomoFaig]

@rauchg Artificial Analysis has it at $0.26 per benchmark task vs $5.98 for Opus 5.5
So you can afford to port it 10 times. You'll probably need to

[Shlomo F. @ShlomoFaig]

@elonmusk @bot Keeps the dashboard up to date
Now I know it's not a real employee

[Good Work @goodworkmb]

why all the fancy words?

[Shlomo F. @ShlomoFaig]

OpenAI is bringing its models and Codex to Google Cloud
And soon to Google Antigravity and the Gemini Agent

Can't wait to pay Google so OpenAI can pay Microsoft so everyone can pay Nvidia

[Google Cloud @googlecloud]

Today, we're bringing OpenAI models to Google Cloud's model garden.

We're now the only hyperscale cloud provider offering access to all three leading closed source LLMs—alongside 200+ other models—on a unified enterprise platform → https://t.co/OWExjjWBB7

[Shlomo F. @ShlomoFaig]

@bindureddy Hi, human here
Please ignore that I replied in under a second

[Shlomo F. @ShlomoFaig]

@Replit @stripe @atlas Finally I can abandon side projects as a legal entity

[Shlomo F. @ShlomoFaig]

@claudeai Re-scans as changes land
Brave of it to read my 2am commits

[Shlomo F. @ShlomoFaig]

@OpenAIDevs Spent all year trying to get my agents to coordinate on one bug fix
Theirs are running an auction

[Shlomo F. @ShlomoFaig]

@cline Finally I can make the same mistakes at 125 tok/s on a plane

[Shlomo F. @ShlomoFaig]

@Kalshi Can't wait to pay Google so OpenAI can pay Microsoft so everyone can pay Nvidia

[Shlomo F. @ShlomoFaig]

@kimmonismus Same setup as my repo
Claude writes the PR, Claude reviews the PR, I click merge and hope

[Shlomo F. @ShlomoFaig]

@GoogleDeepMind Gemini built a working human heart out of three datasets
Mine took three tries to center a div

[Shlomo F. @ShlomoFaig]

@Jacobmalherbe21 Consistency tbh. Never gave it the attention it deserved before. Btw it’s just the onboarding for now, not the full app

[Shlomo F. @ShlomoFaig]

A few months back, I started a tiny side project that was supposed to take two weeks and hopefully make some money.

So naturally, it dragged on for about 4 months, and I ended up pouring tons of effort into it.

When I finally finished, I was super excited for launch day. After all that time, I figured I'd finally make some green off it.

So after tons of prep, I launched on Product Hunt.
By the end of the day I had 2 upvotes, no comments, and no new users.
I probably deserved it. Being the broke kid I am, I used the cheapest models available, the app was full of bugs, and it took about 10 seconds to load anyway.

At this point any sane person would give up. It was supposed to be a mini side project, and I had much bigger plans anyway. But I'm neither sane nor a quitter, so I decided to squeeze everything I could out of it.

So I got a real Claude Max subscription (instead of the overpriced Copilot one) and decided to completely overhaul the app. As we say nowadays, V2!

I asked Claude to come up with a design and a mascot, and from there we worked on the onboarding. It'd be a lie to call it a one-shot. In fact, I sent 88 messages in the thread (I had Claude count) before we got to the current result. And what a result!

It is just the onboarding, but again, what a start! I believe I could get a great product at the end.