RT @yanndubs: 🔥 So excited to share GPT-5!. For thinking mode and API models, we’ve improved performance across key:.- Axes: factuality, st….

Pareto dominance has been achieved externally

This is real and cope at the same time btw.

One interesting thing about the LLM era is we’ve gone from a few model releases a year to a several every quarter. This does weird things to peoples “number go up” expectations.

It has vim keybindings, we're so back.

RT @julieswangg:

RT @sama: not just good at software, good at agentic tasks across the board. also great at long context performance

we only made so many chart crimes because we let 4o make the charts, duh.

RT @jxmnop: most impressive part of GPT-5 is the jump in long-context. how do you even do this? produce some strange long range synthetic….

RT @julieswangg: GPT-5 is here! it's been a wild ride, and i've been honored to work with the dream team (s/o @j_mcgraph @strongduality) to….

RT @SuvanshSanjeev: GPT-5 is what you’ve been waiting for – it defines and extends the cost-intelligence frontier across model sizes today.….

We've open sourced the input links, questions and answers that allow you to construct this same eval at different context lengths up to ~1M tokens. We hope you enjoy it!.

Luckily, we already have the OpenAI BrowseComp eval. It's a great QA dataset with meticulously human graded examples. We turn this into a long context eval by applying hard negative mining to the questions, to find difficult distractors to stuff in the context.

Additionally, these IR datasets have a fair amount of label noise! Manually inspecting the data was a frustrating experience when trying to improve our own models.

Existing long context QA datasets are mostly compiled from old IR datasets. This means you're filling a prompt with lots of random webpages rather than corpora focused on the question. This makes it much easier on the model to ignore a large % of the context.

Along with GPT5, we're open sourcing a new eval, BrowseComp Long Context!. It improves upon existing long context qa evals in data quality and input difficulty. Work with @LK112358, @julieswangg, and our mascot the longham. A bit more below

We are technologically on track for a spectacular century, politics on the other hand.

RT @aidan_mclau: it's pronounced. gee pea toss. enjoy.

everyone into numerology or something today talking about 5s. anyway has me thinking of one of the coolest cars of all time, that came with a flat 5 for some reason

RT @AdrienLE: Gotta hand it to Anthropic, they got to that number more smoothly than we did. (but also check out gpt-oss!) .