Ahti Kitsik

Compression is all you need. CEO & Chief Engineer @arxpoint

[Theo - t3.gg @theo]

Cancelling my Spotify subscription rn

[Bill Gurley @bgurley]

I actually don’t think there is any sign that “model quality” will be the key factor. If anything, the models all seem super close to one another and there is no data point that would point towards someone “gapping” from here. It’s a product race not a model race.

[Ahti Kitsik @ahtik]

@tunguz Highly appreciate you going more nuanced and deep with the posts, thanks a lot!

[Ahti Kitsik @ahtik]

@swyx @cursor_ai @sacrainc I wonder if Claude has been warning about too high demand for weeks due to Cursor IDE's success with Claude 3.5 Sonnet?

Even after the DeepSeek models Sonnet seems to be performing the best.

[Ahti Kitsik @ahtik]

@GaelBreton I'm not so sure about the SEO part. There is plenty of search traffic coming from @perplexity_ai , Claude/@AnthropicAI , etc., which is much better targeted than Google search. There's a good example writeup from the guys at @PlausibleHQ at https://t.co/72UI2rMC2n

[Ahti Kitsik @ahtik]

@iScienceLuvr Not sure how close to your DeepSeek angle, but NVDIA H100 vs Ascend 910C for large scale inference deployment is interesting.

[Ahti Kitsik @ahtik]

@eddychik @tobi @aaronrubin Individual Contributor, people who contribute directly to the project/do the actual work, not managing it.

[Bojan Tunguz @tunguz]

XGBoost is All You Need

What do I really mean by my (in)famous tagline, Part 1

[Ahti Kitsik @ahtik]

@btaylor Have you reflected on how @karpathy's Software 2.0 idea connects to your essay & thoughts? The blog post is 7 years old but even more relevant today than ever. https://t.co/da1QJiQevq

[Ahti Kitsik @ahtik]

@AravSrinivas Whoa. Works great! I wish the upcoming earnings, dividend, and other corporate events date estimate calendar for up to a year in advance would become reliable and stay up to date when changing—it would be extremely valuable.

[Ahti Kitsik @ahtik]

@DiegoT_AI @tiangolo uv is significantly faster than anything else, including Poetry

[Ahti Kitsik @ahtik]

@sheldon_hull @ashtom Sign-up shows still the 3.5 Turbo, I was basing my assumption on that. https://t.co/bsTlcIblt3

[Ahti Kitsik @ahtik]

@skdh Next five plus years will be even more painful, eg Frankfurt airport, Hamburg-Berlin etc trains will be cut off for a few months. Hopefully getting better after that, https://t.co/nShCmqjVWN

[Ahti Kitsik @ahtik]

@Gavmn Thanks a lot for the write-up! I was already very curious about the behind-the-scenes work when the new Linear homepage went live.

[Ahti Kitsik @ahtik]

@robindchnt Using Notebloc happily on both iPhone and Androids. Works great with multipage pdf, corner detection with manual adjustment option, very effective background removal etc. It's 100% free but think I removed the ads at one point for a few $.

[Ahti Kitsik @ahtik]

@simonw "Wrong answers only"

[Ahti Kitsik @ahtik]

@tunguz Curious, is this in the context of NIMs?

[Ahti Kitsik @ahtik]

M2 Ultra with MLX can outperform Nvidia 4090 on inference. Thanks to the unified memory, we are talking about 192GB with M2 Ultra on an Apple Studio for ca 8KUSD. Yes, NVIDIA memory bandwidth is tiny bit faster (1008GB/s for rtx 4090, 960GB/s for A6000 Ada) vs M2 Ultra (800GB/s). MLX is making really fast progress and token per second per dollar performance with 80b open source LLM models is becoming unbeatable ... Will be interesting to see the upcoming M3 Ultra specs ... I wish torch mps would catch up with the performance MLX delivers.

[Ahti Kitsik @ahtik]

@tunguz @nvidia Happy Extreme Gradient Boosting NVersary!