ex yc software eng. currently building: Ness, #1 personal AI Health Coach Atom, coding agent in Rust
Based in United States
- Tweets 169
- Following 23
- Followers 19
- Likes 1,090
I'm happy that the 6-series improved on cost, but saying it's a flatly better model isn't completely accurate
Here's an example of a benchmark where it underperforms compared to the prior versions
69
Apple needs to hire more app reviewers fast. On web you can deploy in minutes, on iOS it takes multiple days, depending on when you get your review.
Increasingly, one bottleneck to shipping quality updates becomes review time..
2
30
would love for you to try Ness, it's an all-in-one AI Health coach with nutrition built in. You can text it photos too if you'd like
happy to send over a lifetime code so you can try it out!
35
Found a fun way to incorporate Jev into Ness!
Whenever you send a message, Jev determines intent and will select an emotion for Ness to display
@typesafeai
2
60
Just launched on Product Hunt! To celebrate, I'm giving away 50 free yearly subscriptions. All you have to do is reply + upvote Ness!
3
1
5
83
Would love to try this out, how is the cost/query so low? Is search data being trained on/sold?
ps. the signup/login screen doesn't seem to render properly on Helium
2
4
819
awesome! i'm working on a harness too, it's very freeing to be able to solve for the annoyances and bloat that exist in the mainstream harnesses
89
the z.ai team might be one of the best in the world in post training if these results hold up, mind-blowing progress
62
Competitive pricing for the lower-end gemini models (3.1 flash-lite, 3.5 flash-lite) would be a plus
currently get outclassed by 5.6 luna and V4 flash 0731 on everything except vision and long context recall
3
343
that said, Fable and Kimi are both relatively token hungry
ideal token efficiency would be 5.6 sol or grok, those are very efficient for their capability
1
19
119k output tokens is for Fable 5 max
fable xhigh, high, medium, and low all produced less output tokens than kimi, with less steps
2
24
The horizontal padding is slightly too much sometimes, here's an ex. vs codex
would be great if padding could be customized in tui.json or similar
3
906
you should try Performant, refreshes take <1s
we do a history computation step which runs in the background and does take some time, but once that completes, refreshes are blazingly fast!
1
2
23
Testing out a fully local version of Performant Chat, running Gemma 4-E2B. One thing that's concerning is moderate/high hallucination rate and tool call error rates.
Should we ship this as an option to Performant?
2
54
