Maxim Saplin articles - 75 total
XYZ% of Code is Now Written by AI... Who Cares?
Microsoft CEO Satya Nadella said that "as much as 30% of the company’s code is now written by...
GPT 4.1, o3, o4-mini - OpenAI releases through the lens of LLM_Chess
This will be a quick post. I've ran the recent OpenAI models through LLM Chess eval: o4-mini and o3...
Mercury Coder - A Quick Test of Diffusion Language Model
I have recently touched on how diffusion/transformer models come into new domains - specifically the...
Llama 4 - 10M Context? Coding? Decent Follow-up?
Meta has brought the long-awaited Llama 4 models on Saturday, April 5. Llama 3 came out on April 26,...
4o Image Gen - Diffusion/Transformer Cross-over Trend?
In March we saw 2 major releases of image generation tools ((Google, OpenAI)) that are very much...
SGLang vs llama.cpp - A Quick Speed Test
Recently, I stumbled upon a post about SGLang, an open-source LLM inference engine that boasts 2-5x...
OpenAI o3-mini Tested in LLM Chess
OpenAI has recently presented its newest reasoning model - o3-mini. At Medium and High reasoning...
Qwen2.5 Max Release Went Unnoticed in Deepseek Hysteria
Last week, Chinese Big-Tech company Alibaba released its best model to date: Qwen-Max. It is a...
DeepSeek-V3: Laziness vs Eagerness
In late 2023, people complained about GPT-4 Turbo's laziness - often the model didn't complete tasks....
Deepseek R1 vs OpenAI o1
Deepseek R1 is out - available via Deepseek API or free Deepseek chat. If you are following LLM/Gen...
OpenAI o1/o3 - Be Careful What you Wish For...
Hallucination is also a latent fear accompanying the copy-pasting of a long scroll of text from a...
OpenAI o3 - Thinking Fast and Slow
OpenAI has teased the o3 model today—a further development of the "reasoning" model and a successor...
Tried Phi-4, It didn't Impress
Phi-4 14B has been recently released. Benchmarks look promising, e.g. it beats GPT-4o in Math: I...
Gemini 2.0 Released, Reminding of "AI Hitting the Wall" Talks
Today Google has presented a major update to its flagship SOTA model - Gemini 2.0. What caught my...
The Power of Pragmatism: Engineering Cultures and China's Ascendancy
A company decided to rename its product XYZ to ABC. Assets, texts in the app, and web pages were...
Can LLMs Play Chess? I've Tested 13 Models (GPT-4o, Claude 3.5, Gemini 1.5 etc.)
UPD January 25, 2025: Deepseek R1 is another model that broke the ceiling of zero wins showing...
Microsoft Autogen Has Split in 2... Wait 3... No, 4 Parts
One of the most popular open-source frameworks of 2023, which explored the application of AI agents...
Llama 3.1 Nemotron 70B - Quirks and Features
The new open model by NVidia, Nemotron 70B has recently been the hotspot of "this is wild", "this is...
DDR5 Speed, CPU and LLM Inference
This is the 3rd part of my investigations of local LLM inference speed. Here're the 1st and 2nd...
Gen AI Hype - the Never Ending Excitement
"When sensationalism wins over nuance, we lose our ability to think." is a great quote by Lex Fridman...
OpenAI o1 Release is so Reminiscent of Apple Events - it's an Incremental Update
Yesterday OpenAI introduced "A new series of reasoning models for solving hard problems", which...
Continue.dev: The Swiss Army Knife That Sometimes Fails to Cut
First Impressions I had low expectations from Continue.dev when I first installed it in...
Python 3.13 RC1 - a Quick CPU Benchmark
Python 3.13 is due to be released in October, yet the first release candidate was published earlier...
llama.cpp: CPU vs GPU, shared VRAM and Inference Speed
This is the 2nd part of my investigations of local LLM inference speed. Here're the 1st and 3rd...
Convergence of LLMs: 2024 Trend Solidified by Llama 3.1 Release
The recent release of Llama 3.1 was reminiscent of many releases this year. It underlined a trend...
DoLa and MT-Bench - A Quick Eval of a new LLM trick
Decoding by Contrasting Layers (DoLa) is a technique suggesting a different approach to calculating...
4090 - ECC ON vs ECC OFF
Fine Tuning LLM via Huggin Face TRL/Torch: ECC On: 2,22 epochs/day ECC Off: 2,33 epochs/day...
MT-Bench: Comparing different LLM Judges
By default, MT-Bench uses OpenAI as a service provider with a gpt-4 model ID, which is a vanilla...
Nvidia's 1000x Performance Boost Claim Verified
Nvidia's keynote at the recent Computex was full of bold marketing and messaging, bordering on...
LLM Fine-tunig on RTX 4090: 90% Performance at 55% Power
At just a fraction of power, 4090 is capable of delivering almost full performance. While running...