Dimitri Zhorzholiani

Co-founder & CEO of | Co-founder & CTO of @Pepzilla_org | Writing about tech, fitness, and innovation.

[Dimitri Zhorzholiani @dimitrizho]

AI passed the Turing Test!

Follow for the latest AI updates and insights: https://t.co/aU2LkVp57G

[Logan Kilpatrick @OfficialLoganK]

Ilya was right and predicted much of this

[Theo - t3.gg @theo]

Man what the fuck

[Dimitri Zhorzholiani @dimitrizho]

@benln 👋🙋

[Dimitri Zhorzholiani @dimitrizho]

The top open models on are now all by Chinese companies

[Dimitri Zhorzholiani @dimitrizho]

GLM-4.6 outperforms claude-4-5-sonnet while being ~8x cheaper.

[Dimitri Zhorzholiani @dimitrizho]

LACT "indirect undervolt & OC" method beats `nvidia-smi -pl 400` on 3090TI FE.

[Dimitri Zhorzholiani @dimitrizho]

OpenAI usage breakdown released

[Dimitri Zhorzholiani @dimitrizho]

Unsloth Dynamic GGUFs - Aider Polyglot Benchmarks

[Dimitri Zhorzholiani @dimitrizho]

GLM-4.5 is now leading the Berkeley Function-Calling Leaderboard V4, Beating Opus 4

[Dimitri Zhorzholiani @dimitrizho]

The "speedup", is only at large context. they demonstrate lack of loss of performance, not gain of performance.

They developed framework to replace in optimal way subset of attention layers in existing models with linear layers, as this suibset in non-trivial to find, and just slapping linear every say 3 layers instead of attention will degrade performance.

[Dimitri Zhorzholiani @dimitrizho]

LLM speedup breakthrough? 53x faster generation and 6x prefilling from NVIDIA

[Dimitri Zhorzholiani @dimitrizho]

Qwen Wan2.2-S2V is coming soon

Alibaba is doing his Alibaba things!

[Dimitri Zhorzholiani @dimitrizho]

OpenAI, I don't feel SAFE ENOUGH

[Dimitri Zhorzholiani @dimitrizho]

Their paper, “Inverse Scaling in Test-Time Compute,” reveals a surprising phenomenon: in certain tasks, models like Claude and OpenAI's GPT-o series actually perform worse when allowed to "reason" for longer.

[Dimitri Zhorzholiani @dimitrizho]

Anthropic’s New Research: Giving AI More "Thinking Time" Can Actually Make It Worse

[Dimitri Zhorzholiani @dimitrizho]

Performance is insane

[Dimitri Zhorzholiani @dimitrizho]

Qwen3-Coder is released! This 480B-parameter Mixture-of-Experts model (35B active) natively supports 256K context and scales to 1M context with extrapolation

[Dimitri Zhorzholiani @dimitrizho]

UTCP: A safer, scalable tool-calling alternative to MCP

[Dimitri Zhorzholiani @dimitrizho]

Kimi K2 tops creative writing benchmark

[Dimitri Zhorzholiani @dimitrizho]

Moonshot AI just made their moonshot