Tim Kellogg

i still find it hilarious that there’s no number after it. there’s not going to be a number 2

one way it might help — being ClosedAI might be hurting the research. By opening up a model that shares common critical aspects, researchers can collaborate outside of OpenAI on ideas

what benefits are there to open source AI though? it costs a ton, if you get roasted your entire lab loses credibility, and nobody can really give back without a crap ton of capital

grand ~= sounding somewhat sarcastic

he’s a SV CEO, he doesn’t do things out of the goodness of his heart. he makes product. what product play was this?

flowchart TD subgraph Home user gpt-oss end user --> gpt5 --> web gpt5 --> gpt-oss --> MCPtools

one criticism i haven’t heard: gpt-oss is a bad agent

they say lots of bad stuff about it, but its tool calling and long-horizon reasoning seems on point

what if this is the plan — gpt-oss on the edge, only to agentify local resources

yes i don’t believe for a second that sama didn’t have some grand strategy with the open weights

i’m still thinking it’s going to be an edge piece somehow. maybe on a phone/laptop or his new affair with Jonny Ive

lol, they’re super smart guys though

yeah, simply basing on GPT-4.1 is going to be a big deal. i’m curious to what extent it’ll actually route

that’s my boi

Qwen/Qwen3-4B-Thinking-2507 · Hugging Face We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co/Qwen/Qwen3-4...

OpenAl v @OpenAI Follow LIVETREAM THURSDAY 10AM PT

GPT-5

Bar chart comparing performance of four Qwen3-4B model variants—Thinking-2507 (red), Thinking (gray), Instruct-2507 (blue), and Non-Thinking (beige)—across five benchmarks: 1. GPQA • Thinking-2507: 65.8 • Thinking: 62.0 • Instruct-2507: 55.9 • Non-Thinking: 41.7 2. AIME25 • Thinking-2507: 81.3 • Thinking: 65.6 • Instruct-2507: 47.4 • Non-Thinking: 19.1 3. LiveCodeBench v6 (25.02–25.05) • Thinking-2507: 55.2 • Thinking: 48.4 • Instruct-2507: 35.1 • Non-Thinking: 26.4 4. Arena-Hard v2 • Thinking-2507: 34.9 • Instruct-2507: 43.4 • Thinking: 13.7 • Non-Thinking: 9.5 5. BFCL-v3 • Thinking-2507: 71.2 • Thinking: 65.9 • Instruct-2507: 61.9 • Non-Thinking: 57.6 Across all benchmarks, the Thinking-2507 variant (red bars) consistently leads in performance except for Arena-Hard v2, where Instruct-2507 scores highest. A watermark-style logo is faintly visible in the background behind the bars.

Qwen3-4B Instruct & Thinking

uuuh, guys this isn’t a boring model

This crushes all the agentic benchmarks, even beating out the already-impressive qwen3-30b-a4b

It’s hanging with some already impressive mid-sized models at only 4B

a close up of a man 's face with a blurred background ALT: a close up of a man 's face with a blurred background

some of their marketing videos are eye opening

Trump says it’ll be 3000% cheaper

imo that’s a failure mode — i’ve been fairly successful at staying high level, checking in every 20 min, and using the main time to do meetings, docs, etc.

if i have to step in and actually write code it jams up my rhythm, and my day basically halts. so i end up doing a lot of up-front planning

yes!

word is GPT-5 will have a 13.5% increase in version number over GPT-4.5

Junyang Lin • &7 @JustinLin610 drinking still. will announce later. 4b is fabulous. i do love it. 11:53 AM • 8/6/25 • 4.2K Views

qwen coming in hot in their usual style

just sent a meeting request with “..unless GPT-5 comes out” in the title

i’m not sure you’re right about “that ship has sailed”

it’s just that there’s tiers of openness, more is better than less, but the tiers aren’t equidistant

yes, that’s what i’m talking about!

Okay, I'll chime in, as a person with less mental health than average.

Working with agents can be really easy or really trying. It's really easy when you remember that the agent won't get mad at you or walk out of the meeting. It's easy when you remember that agents aren't scary like people.

you should

interestingly, i got so much flak that i wrote a retraction, and the retraction was far more popular than the actual piece. idk…

signal 1: i’m working with a marketing person at work who’s fucking killing it with claude code

signal 2: some people are discovering that frameworks are often not worth the time anymore

i think both are still panning out

Normware: The Decline of Software Engineering

i’m getting some early signals that i may have been right about this, i wrote back in January

tl;dr is software wont be boxed & shipped unless it solves a truly hard problem

timkellogg.me/blog/2025/01...

honestly i just need this but for myself

bsky.app/profile/timk...

they overdial on speed. @cameron.pfiffer.org noticed that they can’t tool call for shit

the crazy part is they use their own special quantization. while it’s also 4 bits, it’s a different scheme that what openai trained with, so it’s braindead in subtle ways and they don’t care bc it’s fast

alright, let’s see if my magic dice work today

there will be releases today but not the cool stuff