Pinned Tweet
We (Emochi, with 10M+ users) stopped using benchmark scores for production decisions about 6 months ago.
The pattern was consistent: models ranking top on benchmarks underperformed in real usage. Models ranked lower drove better retention.
A few notes on what we learned:
6
9
1,029
“Everything we set out to do this quarter, we did. In most cases, we did more.”
– Arkady Volozh, founder and CEO
Highlights from our Q2 results:
Total contract value won during the quarter up 4x;
Four landmark customer agreements averaging over $1 billion in TCV;
Average annual contract value above $20 million per MW;
AI cloud revenue up 514% to $575M, and a higher adjusted EBITDA margin of 50%;
Contracted capacity guidance for 2026 increased to 5 GW;
Expecting to deploy more than 1 GW per year from 2027.
Sharing some thoughts on AI evals for consumer AI
Why serious AI builders are skipping third-party evals techradar.com/pro/why-seriou…
10 million users.
100+ models tested every week.
More than 1 trillion tokens processed every day.
Henry Wang explains why he thinks the next generation of consumer AI won't be won by the best model, but rather by the teams that can iterate and scale the fastest.
Listen now: nebius.com/podcast
@Henry_Flowgpt @Josh_Liss
I built a @claudeai skill to rate AI proficiency.
I think Paxel (paxel.ycombinator.com) is a good start, but it doesn't work for me because of how sensitive my data is — it uploads your code to servers
So I made an open sourced one. Repo in the first comment
2
4
253
Some recent thoughts on company efficiency linkedin.com/pulse/closing-l…
2
109
Two things worth putting time developing agent in production
1. Providing accurate context
2. Writing good test cases
127
Henry Wang retweeted
Open-sourced Claude Code Web.
For people who can’t run Claude Code Client locally, or prefer a browser workspace over the TUI.
Runs with local Claude Code CLI or on a remote server via SSH tunnel.
Most importantly: you can change it into any skin you like.
Open-sourced Claude Code Web.
For people who can’t run Claude Code Client locally, or prefer a browser workspace over the TUI.
Runs with local Claude Code CLI or on a remote server via SSH tunnel.
Most importantly: you can change it into any skin you like.
One takeaway for 2026: Eval infrastructure will determine product ceiling more than model selection.
The question isn't "which model is best." It's "which system learns fastest from real users."
Wrote up the full framework: kaon.io/blog/offline-evaluat…
3
230
We rebuilt the pipeline on three components:
1. Elo/TrueSkill on real conversations — filters out bad models in hours
2. Conversation-level A/B — not user-level (key distinction below)
3. Reward models trained on behavioral signals — closes the loop into training
2
187
Core issue: benchmarks optimize for "correct responses" on discrete tasks.
Consumer AI optimizes for "experiences that feel right" across continuous interactions. These objectives aren't aligned. At scale, they actively conflict.
1
170
Henry Wang retweeted
Interviews are broken. resumes mislead.
We helped 100k+ people land jobs & scaled to $10M ARR. now we’re rebuilding hiring from scratch.
meet WorkTrial AI — where companies see the real work before they hire.
If your product relies on user-facing prompt features, this post is for you. I’m sharing hard-won lessons from 10,000+ prompt iterations across complex, structure-sensitive workflows where every percentage point of success rate mattered.
linkedin.com/pulse/first-pro…













