LIVEReading: NewsUpdated: 10 min agoSubscribers: 23,400 LIVEReading: NewsUpdated: 10 min agoSubscribers: 23,400
CBW

OpenAI's GPT-Red teaches itself to be harder to break, NVIDIA tops the retrieval charts

OpenAI published GPT-Red, a model that improves its own robustness through self-play. NVIDIA's Nemotron 3 Embed just hit #1 on the retrieval benchmark RTEB.

OpenAI published research on GPT-Red, a model that uses self-improvement techniques to become more robust — essentially finding its own weaknesses and patching them. If this approach scales, it changes how safety and reliability work for builders who depend on GPT models in production.

New models

GPT-Red: self-improvement for robustness

OpenAI's GPT-Red paper describes a model that generates adversarial inputs against itself, then trains on the failures. The goal is a model that holds up better under edge cases and deliberate misuse. This is research-stage work — there is no GPT-Red product you can call via API today. But the direction matters: if self-improvement loops become standard, model reliability could improve without waiting for a full new training run.

NVIDIA Nemotron 3 Embed hits #1 on RTEB

NVIDIA's Nemotron 3 Embed model ranked first overall on RTEB (Retrieval Text Embedding Benchmark), which specifically tests how well embeddings hold up in agentic retrieval tasks — the kind of multi-step search an AI agent does when it needs to find the right chunk of a document before answering. The model is available on Hugging Face. If you are building a RAG pipeline or document search tool, this is worth benchmarking against whatever embedding model you are using now.

Soofi S: open 30B model, strong in English and German

A German AI consortium released Soofi S, a 30B open model that reportedly tops benchmarks in both English and German. That bilingual performance is rare at this size. If you are building anything for European users — customer support, document processing, content tools — Soofi S is worth a look. It is open, so you can run it yourself.

Open-source releases

Gemma 4 26B on a 13-year-old server with no GPU

A Hacker News post with 299 upvotes walked through running Google's Gemma 4 26B at 5 tokens per second on an old Intel Xeon with zero GPU. That is slow but usable for batch jobs. The practical takeaway: you do not need a GPU server to run a capable 26B model. Old hardware you already own might be enough for offline document processing or overnight batch tasks.

Axolotl v0.18.0 and browser-use 0.13.6

Axolotl, the fine-tuning framework, shipped v0.18.0. Browser-use, the library that lets AI agents control a browser, released 0.13.6. Both are incremental updates. If you are already using either tool, update. If you have not tried browser-use, it is one of the more practical ways to give an AI agent the ability to fill out forms or scrape pages without writing custom Playwright code.

Research worth reading

The Little Book of Reinforcement Learning

A free, open GitHub book on reinforcement learning picked up 168 upvotes on Hacker News. It is aimed at people who want to understand how RL actually works — relevant now that RL is behind most of the reasoning improvements in models like o3 and Gemini 2.5. Not a tool, but a useful read if you want to understand why models behave the way they do.

What builders can do this week

1. Swap your RAG pipeline's embedding model for NVIDIA Nemotron 3 Embed and run a quick retrieval accuracy test on your own documents. It is on Hugging Face and free to download.

2. If you have an old desktop or server sitting unused, follow the Gemma 4 26B Xeon guide and set it up as a local inference box for overnight batch summarization or classification jobs — no cloud costs.

3. Install browser-use 0.13.6 and build a small agent that logs into one of your own web tools, pulls a report, and saves it to a file. It is a two-hour project that shows you exactly where browser agents break.

Honest note

// what we actually tested

Honest note

Confirmed: OpenAI published the GPT-Red research post at the URL listed. This is a research paper, not a product launch — no API access announced.

Confirmed: NVIDIA Nemotron 3 Embed is available on Hugging Face and the RTEB #1 claim comes from NVIDIA's own Hugging Face blog post. CBW has not independently run the benchmark.

Not independently verified by CBW: Soofi S benchmark claims come from a single The Decoder report. We have not tested the model or reviewed the benchmark methodology.

Not independently verified by CBW: The Gemma 4 26B on Xeon performance figures (5 tokens/sec) come from a single blog post. Results will vary significantly by hardware configuration.

Worth noting: The OpenAI Cars24 and teen AI access posts in today's signals are marketing and policy content, not product news — we skipped them.

Source: OpenAI GPT-Red research post — https://openai.com/index/unlocking-self-improvement-gpt-red

Source: NVIDIA Nemotron 3 Embed — Hugging Face blog — https://huggingface.co/blog/nvidia/nemotron-3-embed-wins-rteb

Source: Gemma 4 26B on a 13-year-old Xeon — Neomind Labs — https://www.neomindlabs.com/2026/06/08/running-gemma-4-26b-at-5-tokens-sec-on-a-13-year-old-xeon-with-no-gpu/

Source: Soofi S open 30B model — The Decoder — https://the-decoder.com/german-ai-consortium-releases-soofi-s-an-open-30b-model-that-tops-benchmarks-in-both-english-and-german/

Source: The Little Book of Reinforcement Learning — GitHub — https://github.com/alxndrTL/little-book-rl/

Source: browser-use v0.13.6 — GitHub — https://github.com/browser-use/browser-use

// daily build

One project. 5 minutes. Daily.

Get tomorrow's best AI project in your email. With a guide that works. Free. No spam.

23,400 builders read this