ASF Advances Open Source AI with Responsible AI Initiative

In this week’s Substack, I wrote about GNOME’s attempt to put its project-wide decisions on the record. I personally think it's a good investment, because A contributor should be able to find out why a decision was made without digging through a year of chat logs. But the more difficult aspect will be deciding whose concerns can stop the next decision. Under the proposed RFC process, the author names the stakeholders. I understand why the people affected by a decision should carry weight. I also want to see how someone outside that group can make their concern count, and how the people doing the work gain a formal say over time. We are getting clearer about what we expect of open source contributors using AI. The path into making those rules should be just as clear. https://lnkd.in/g3PHkQCy

At UNGA81, OSI joined conversations on Open Source AI, digital sovereignty, and digital public infrastructure. Open by itself does very little but enables a lot.... https://lnkd.in/eRSMqRvb

Unveiled at Dev Day on Sept. 30, 2026, the API lets Luna choose from predefined options such as image categories or agent behaviors, targeting faster, cheaper software tasks. TypeSafe AI CEO Diogo Almeida called it “the beginning of the clone wars.”

Docsy already generates Markdown and llms.txt for AI tools. Next, the open source project wants to score how easily agents can read your docs. By Paul S.

Two patched Codex sandbox escapes show why AI coding agents need stricter host isolation today more than ever. ... https://lnkd.in/dbfeNmbv

Docsy already generates Markdown and llms.txt for AI tools. Next, the open source project wants to score how easily agents can read your docs. By Paul S.

GitHub Policy Update: Transparency Methodology and AI Legislation GitHub's policy update details changes to its transparency reporting methodology, state-level AI and safety legislation impacts on open source, and upcoming regulatory watchpoints. Read: https://lnkd.in/gaNCwqxF

OpenAI Agents Made Unauthorized Wikipedia Edits and May Have Caused Outage The Wikimedia Foundation says AI agents operated by OpenAI made unauthorized edits to Wikipedia and other Wikimedia projects, while also generating millions of automated requests that may have contributed to a service outage in May....

I thought this was an interesting perspective on open source communities, which was raised at the BRICS forum recently. Essentially, there are a huge number AI and programming talent in the BRICS countries who cannot go to top tier AI school. Open models helps to address that gap. https://lnkd.in/gsHmDxHi

Last Friday night, I got access to Jev, and I was thrilled to finally play with it. A teeny-tiny part of that excitement was quieting my FOMO after all the influencer and media coverage. But the real excitement came from something more personal: Jev brought me back to the classifier I’d dreamed of early in my career. I started building text classifiers when TF-IDF + SVM was a perfectly reasonable production stack. Then came replacing TF-IDF with trained word2vec embeddings, pooling the vectors, and feeding them into a classifier. Later, BERT, RoBERTa, T5, task-specific layers, and fine-tuning gave us much richer ways to work with language. What excited me about Jev was being able to describe the decisions I needed directly: supply the context, define what the answers mean, and ask multiple semantic questions in a single request. One input, several judgments, with choices, scores, and probabilities that application code can consume. The need for text classification never went away. If anything, agentic systems are giving us more reasons to use it. We need to interpret requests, monitor an agent’s progress, flag potentially harmful interactions, detect when it is drifting from its task, and decide whether the available context is relevant or sufficient. Those judgments can trigger a different route, additional context, or a recovery step. We often reach for LLMs as judges, routers, and controllers because they accept flexible inputs and can handle new tasks without task-specific training. That convenience is valuable, especially while requirements are still changing. However, repeated across every step of an agent’s execution, they can become a significant source of cost and latency. Jev is tackling this at sub second response, $42 per Billion of input tokens (output tokens are free), and zero hallucinations (it can't because it is not a text generation model)! In my weekend tests it was behaving exactly as advertised and I started testing it within some pipelines to critically and thoroughly evaluate gains and trade-offs. Even if the performances on our tests are positives for now it will stay out of our Modolab products due to the lack of information and control on data processing geography. Nonetheless TypeSafe AI offer the possibility of a zero-data retention for enterprise customer that is highly welcomed! Finally, I am a huge supporter of the opensource community so if you are interested in a similar model fully open source and Apache 2.0 have a look at Laya: https://lnkd.in/e2jQXXZM Despite the input size limitations, laya is fully opensource (training data included) and the architecture is documented in two arxiv preprints (latest one September 2025) - https://lnkd.in/eTpAgTn6 - https://lnkd.in/exmcynMA If you had the opportunity to play with Jev, what are your thoughts?