Product / AI agent builder. 12 years shipping to production. Google → Microsoft → Startups. now deploying agents in businesses.
London, England
Joined February 2015
- Tweets1.1K
- Following1.1K
- Followers319
- Likes2.4K
Pinned Tweet
My goal on X: 1,000 genuine connections with people actually building AI.
Looking to connect with:
AI agent builders
Engineers shipping to production
Founders making AI work in their business
Anyone who's had an agent break at 4pm on a Friday
If that's you, say hi below 👇
Google released Gemini 4 Argon today.
The benchmark numbers are whatever.
This is the bit I care about:
They gave it a 205k-line production codebase, one goal, and left it running overnight.
It came back 11 hours later with 6 proposed fixes.
Humans rejected 4.
2 shipped to production.
One cut long-session overhead by 40%.
That feels like a much better test of coding agents than another “build a Todo app” demo.
It’s only Wednesday, but 3 AI shifts this week:
1. Mistral + Reflection are pushing strong open models, meaning more companies may be able to run AI themselves, keep data in-house and pay less.
2. AI watermarking is getting real. Google now lets anyone check supported images, video and audio for hidden AI marks.
3. “Decision models” are becoming a thing: smaller models whose whole job is deciding what an agent should do next.
Google just released a tool that lets anyone check whether an image, video or audio file carries an AI watermark.
How I think works is actually pretty simple:
Some AI systems hide a tiny invisible pattern inside the stuff they generate. You can’t see or hear it, but the detector can.
It's like a little digital stamp saying “this came from an AI system."
Obviously it only works if the AI tool added that stamp in the first place..
lInk in the first comment
Perplexity just released an open “decision model".
Give it some context + a set of options and it gives you probabilities instead of writing a paragraph.
Tiny boring decisions are going to be a surprisingly big part of making agents actually work.
Are we moving from “one big model does everything” to smaller models whose whole job is just to make good decisions?
pplx-decider-v1.1-27b, our updated open weights multimodal decision model, is now available.
It scores the highest on the new @huggingface Decision Index 0.3 benchmark. pplx-decider-v1.1-27b costs half as much as v1, at $0.02 per million input tokens.
huggingface.co/spaces/multim…
Google Docs just became a little more agent-native.
You can now open, edit and collaborate on .md files directly in Docs/Drive without converting them.
Sounds minor, but Markdown is basically the working language of coding agents.
A useful little bridge between humans working in Docs and agents working in files.
This caught my attention more than the benchmark chart.
Reflection's new Beam model has 501B parameters, but only 23B active at a time and they claim performance around the Chinese open-model frontier.
That matters because Qwen/GLM/DeepSeek have increasingly owned the open-weight conversation.
A serious alternative would be useful, but with one caveat: the weights aren't actually out yet and these are still their own benchmarks.
So, not proven yet, but interesting.
Mistral just dropped Large 4.
1 trillion parameters, but only 49B active at a time.
Their claim: best open-weight model from the US or Europe, with particularly strong results in cyber, finance and manufacturing.
Basically Europe saying: we’re still in this race.
Interesting little shift in AI coding.
Lovable/Replit already let you build + ship inside their world.
I came across a tool taking a different approach: stay in Codex, Claude Code, Cursor etc and let your existing agent handle hosting, database, domain, auth, email and analytics.
The interesting bit to me isn't deployment.
It's that the infrastructure layer is becoming independent of whichever agent you use to build.
AI this week:
• AI started writing jailbreak instructions to itself
• AI actors started randomly speaking Cantonese
• AI agents can now phone people
Anyway how was your week?
השארתי את זה מנגן.עלI שירים שלא שמעתי שנים עלה מלאכי חזקיה. הגיעה התוכנית הבאה.ולרגע זה הפסיק להרגיש כמו מערכת שבניתי. זה פשוט הרגיש כאילו הרדיו של אז חזר לעבוד.