Florian Dedov

🧠 AI/ML Engineer
📹 YouTube: GitHub:

[Florian Dedov @NeuralNine]

I was able to meet Jensen Huang the CEO of @nvidia at #GTCParis.

This guy is a legend!

[Florian Dedov @NeuralNine]

Claude models repeatedly ignore my instructions in prompts and in my CLAUDE.md file.

It also seems to get worse over time...

[Florian Dedov @NeuralNine]

@ThePrimeagen "I nuked your files with sed"

[the tiny corp @__tinygrad__]

This is a great sustainable business model for open weights. It's free if you run it yourself, but if you are running a cloud providing it to others for money, you should have to share profits. (from Kimi K3 License)

[Florian Dedov @NeuralNine]

Everyone loving Kimi K3.

Anthropic meanwhile:

[Kimi.ai @Kimi_Moonshot]

Releasing the model weights and technical report of Kimi K3.

Kimi K3 is our most capable model: a 2.8T MoE model with native visual understanding and a 1M-token context window.

New model architecture: 2.5x the intelligence per unit of compute, not just more params.

Alongside Kimi K3, we're opening up more of the stack behind it — high-performance attention kernels, MoE communication library, and infrastructure for running agent environments at scale.

Model weights: https://t.co/7m7eEg6Y0B
Tech report: https://t.co/yeu6cjpMCT
Tech blog: https://t.co/YTfiMSNM1f

[Naval @naval]

When software was expensive - thin, horizontal, best-of-breed software stacks extracted rents across every business.

Now that software is cheap - value moves to vertically integrated businesses that deliver opinionated end-to-end experiences.

[Florian Dedov @NeuralNine]

Imagine you copy someones homework and then run to the teacher to snitch on someone copying it from you...

[Zain @zainhas]

Kimi K3 on legal tasks is ~2x fable perf 🤯

The benchmark is from Harvey and involves hard autonomous legal work

Kimi K3 at 26.7%, vs Claude Fable 5 at 14.2%.

[Florian Dedov @NeuralNine]

This is so fun...

[Florian Dedov @NeuralNine]

Codex = Windows
Claude Code = macOS
OpenCode = Ubuntu
Pi = Arch Linux

[Florian Dedov @NeuralNine]

If there was a benchmark for over-engineering all OpenAI models would rank at the very top...

[Florian Dedov @NeuralNine]

@theDrewDag I hate both but Claude is better.

[Florian Dedov @NeuralNine]

Every time I run Codex I am just disappointed. Why is it so popular?

[Florian Dedov @NeuralNine]

I started preparing a video on the state of open weight models and how it takes 6-12 months for them to catch up with the state-of-the-art.

I'm glad I waited...

[Florian Dedov @NeuralNine]

Looks like Anthropic will have to start distilling Kimi K3.

[Arena.ai @arena]

Big news: Kimi-K3 by @Kimi_Moonshot is now #1 in the Frontend Code Arena with 1679 pts, surpassing Claude Fable 5.

This is a 17-place jump from Kimi-k2.6 (#18 -> #1).

In Frontend, Kimi-K3 ranked #1 in 6 of 7 domains: Brand & Marketing, Reference-Based Design, Data & Analytics, Consumer Product, Simulations, and Content Creation Tools, landing #2 only in Gaming behind Fable 5.

The full model weights will be released by July 27.

Congrats to the @Kimi_Moonshot team on this major milestone!

[Florian Dedov @NeuralNine]

@mitsuhiko Would love to see some of them coming out of Europe as well though.

[Florian Dedov @NeuralNine]

Kimi K3 on Opencode Go is abysmally slow...

[Florian Dedov @NeuralNine]

Is Kimi K3 worth the price increase compared to GLM 5.2?

[Florian Dedov @NeuralNine]

I'm surprised by how fast and capable GLM 5.2 Max is using the Pi harness.

[Florian Dedov @NeuralNine]

@theDrewDag Thanks

[Florian Dedov @NeuralNine]

What's the best harness to run GLM 5.2 in?