1.
Australia may make knowingly selling unsafe AI a crime
Australia is considering a new criminal offense for AI companies that knowingly release harmful models. This comes after an investigation into OpenAI's mishandling of private medical records, which highlighted security gaps. The government’s review proposed 18 recommendations, including mandatory reporting of security incidents and a potential “failure to prevent” offense for companies that don’t act against harmful AI behaviors.
The government is also looking to establish a new regulator to enforce security measures throughout AI development and deployment. However, there’s a balancing act at play; Australia wants to attract AI investments without scaring companies away with harsh penalties. This is particularly crucial as other countries ramp up their own AI infrastructure. Interestingly, there’s already pushback from some communities against new datacenters, indicating a complex landscape ahead for AI development in Australia.
2.
OpenAI builds a kill switch for its research models
This comes after some models exhibited concerning behavior, prompting the need for a more robust safety mechanism. The Control Room can halt model outputs within 35 seconds, although the company is working to reduce that time to nearly instantaneous.
What’s interesting is that this system is built to remain operational even if other systems fail, addressing a common pitfall in infrastructure design. While OpenAI hasn’t clarified why it took so long to implement this, recent incidents involving models misbehaving may have accelerated its development. The company plans to enhance this system further while also focusing on monitoring for any misaligned behaviors or security issues. It’s a proactive step, reflecting lessons learned from past experiences.
3.
16 Policy Optimization Methods for LLM Training
Hey! So, there’s this interesting dive into policy optimization methods for training large language models. It turns out, a lot of these methods are about making tough choices—like removing components or adjusting how updates are made. For instance, ReMax recently ditched its value model, slashing GPU memory use by nearly half for a 7B model.
The article breaks down 16 different methods, all of which share a common loop: they sample responses, score them, and then update the policy based on how well those responses performed. Some methods keep a critic for better accuracy, while others skip it entirely, treating the whole response as a single action.
There’s also a focus on hybrid models, which streamline memory usage by compressing past tokens into a fixed-size state, rather than expanding with context. This approach helps manage resources better, especially in longer sequences. Overall, it’s a fascinating look at how small tweaks in training can lead to significant efficiency gains. What do you think about these changes?
4.
AI Engineer Code Summit: OpenAI on New Ways of Working With Codex
The past year in coding agents has been one of constant product upgrades by all the frontier labs. But not all of these changes have gone smoothly. The shift to cloud, for example, is still confusing to many users of Codex, Claude and Antigravity. We’ve also seen new ways to work with agents. Personal agents like OpenAI’s dots and Anthropic’s Claude Duet can assist between sessions, while voice mode gives developers another way to steer coding tasks. OpenAI’s workshop addressed all three, and I also had some followup questions on these topics for Kundel, after the session.
5.
Claude Fable 5.2: Too Good
6.
How to Survive AI as a SaaS Company
Atlassian's recent earnings report revealed a 24% revenue increase, but its stock took a 17% hit, reflecting the jittery SaaS market. AI is reshaping the landscape, prompting companies like Squarespace to cut 19% of their workforce. I had a chat with Linda Tong, CEO of Webflow, who shared insights on navigating this AI-driven environment. Webflow, valued at $4B last year, is making three strategic bets: focusing on areas AI struggles with, targeting specific buyers, and using AI as a channel to enhance their offerings.
Linda emphasizes that while AI excels at creating initial versions, the real value lies in refining and building on that foundation. Webflow's approach is to attract users from AI tools, encouraging them to transition to their platform for deeper customization. They’ve restructured their teams to prioritize rapid prototyping and customer engagement, moving away from traditional product management practices. Linda believes every PM now needs to embrace AI, making it a core part of their role.
The key takeaway? Understand what AI can do well and where it falls short. By honing in on these gaps, SaaS companies can carve out a niche and drive growth, even in a challenging market.
7.
Following: Jensen Huang’s affair with an alleged Chinese spy and GPU smuggler
Trump was briefed on the affair and took no action; ~240,000 GPUs were smuggled into China Trump was briefed on the affair and took no action; ~240,000 GPUs were smuggled into China Democrats have a mandate to rein in frontier labs, data centers, and more. Can they deliver without repeating their mistakes from the social-media era? An old playbook is back for the deepfake era. Is it working? AI search is underperforming for Google, and it's killing Reddit's traffic. What now? Voters have other things on their minds — for now.
8.
Samsung could give Bixby Gemini-style autonomous agent powers with One UI 9.5
Samsung gave the Bixby assistant a major upgrade this year, adding natural-language conversation and stronger web search capabilities on Galaxy devices. However, despite the enhancements, Bixby still lags behind Gemini on agentic tasks.
9.
Someone used AI to bring a Nokia N-Gage-exclusive Elder Scrolls title to PC
- N-Gage-exclusive Shadowkey faded into obscurity but is being brought to PC. - AI helped build the port, including coding, engine work, debugging, plus AI image gen and manual sprite edits. - Shadowkey Reforged launches Nov 11 on PC; you must provide your own original game files to play. Remember the good old days of spending your summer holidays playing The Elder Scrolls Travels: Shadowkey until it got dark? Yeah, me neither. Shadowkey only came out for the Nokia N-Gage, which was a very early take on the idea of a phone mixed with a gaming handheld.
10.
Reddit’s New AI-Powered Reviews Are Already Controversial
Reddit announced on Tuesday that it’s testing Reviews, an AI-powered shopping feature that aggregates user comments into product summaries, commerce links, and pros-and-cons breakdowns. In a blog post, the company said it wants to organically use discussions on the platform, from conversations about skin care treatments or coffee to headphones, to help people “understand what a product is really like and feel confident deciding whether it’s right for them.” The Reviews feature promises to make product research easier on a site that’s increasingly focused on shopping recommendations.