Reading Room

A video-editing startup called invideo says the new OpenAI model GPT-6 Astra lets it plan edits with more precision, triples the accuracy of its color correction and grading, and generates 50 custom visual effects in a single day. A concrete, working example of what OpenAI's newest model can actually do in production, not just a demo claim.

A new open-source project called Jeff gives you a tiny, fast AI model you can drop straight into your own code whenever you just need a quick yes-or-no or pick-one-of-several decision, without paying for a full model call, a building block that fits directly into agent pipelines like the ones you run. Jeff is a set of small models, retrained versions of two existing base models called Qwen3.5 and Gemma 4, built to sort things into categories they were never specifically trained on, just described in plain words at request time (zero-shot classification). You describe a situation in plain English, list the possible answers, and Jeff hands back a probability for each option after a single pass, with no generated text to parse. It answers in about 22 milliseconds on a high-end workstation graphics chip (an RTX PRO 6000) and about 28 milliseconds on an Apple M4 Max Mac chip, using Apple's MLX system for running AI models. It can also answer several unrelated questions about the same piece of text in one request, useful if a pipeline needs several classifications on one input at once.

Jeff is built to run on ordinary hardware end to end. The smaller, 0.8-billion-parameter version (parameters are a rough measure of model size) trains in about two hours, and the larger 2-billion-parameter version in about three and a half hours, both on one RTX PRO 6000 chip. All the training examples were written by an open model called Qwen3.8-Flash-Next running on two small Nvidia AI computers called DGX Sparks. Testing happened on an ordinary MacBook. Jeff uses the same request format as an existing commercial product called Jev, but it is a separate, independent project, not affiliated with Jev's maker, a company called TypeSafe; its training method builds on an open-source recipe called AutoJev. Three finished models are posted publicly on Hugging Face, a site that hosts AI models for download: Jeff-Qwen3.5-0.8B, Jeff-Qwen3.5-2B, and Jeff-Gemma4-E2B.

The team tested Jeff on 4,599 questions across five standard tests everyone runs, plus a separate 105-question hard set from JevBench. Untrained, the small Qwen3.5-0.8B model scored 45.3 out of 100 overall; after Jeff's retraining, the same-size model scored 79.1, and the larger Jeff-Qwen3.5-2B scored 83.1. The commercial Jev model scores 83.0 on the same kind of test, and a much bigger, 27-billion-parameter open model called AutoJev-27B scores 84.9. In plain terms, a model more than thirty times smaller than that 27-billion-parameter model lands in the same range on everyday classification and fact-checking tasks. On one test measuring whether an answer sticks to its source material (RAGTruth), Jeff-Qwen3.5-2B scored 88.9, tying the far larger AutoJev-27B. On Financial PhraseBank, a test of reading financial statements, the untrained small Qwen3.5-0.8B model scored just 36.0, and after Jeff's retraining it jumped to 96.4. On WinoGrande, a common-sense reasoning test, Jeff-Qwen3.5-2B rose from 52.2 untrained to 79.0 after retraining. The gap does show up on harder reasoning tests: on three tougher tests, BBH, JudgeBench, and JevBench's hard tier, Jeff stays well behind the bigger models, which the makers say is expected at this size.

To see how Jeff handles situations nothing like its training data, the team had it play three video games, Doom, Frogger, and Pac-Man, choosing each move from a plain-language description of the options and their consequences, never told which move was correct. In Doom, the small Jeff-Qwen3.5-0.8B model averaged 6.55 kills per run, matching a hand-coded rule-based bot and far ahead of the same untrained base model's 5.0. It answered in 29 to 49 milliseconds per move on an M4 Max Mac, versus the 212 milliseconds per move that Jev's own published Doom test took calling an outside service.

A LinkedIn roundup counts 450 million repositories on GitHub and lists the 20 most-starred AI projects right now, including the AI coding agent OpenClaw at 278,000 stars and Anthropic's own Claude Code at 75,000 stars. A quick way to see which open-source AI tools are actually gaining real traction, useful for picking what to try next.

Five Business Insider staffers tried putting AI agents to work on real tasks and share exactly what worked and what backfired. You are running your own fleet of AI agents, so firsthand accounts of real failure modes are worth a look before assuming your own setup is solid.