ChatGPT remembers things you never asked it to
Researchers analyzing chat logs from 80 ChatGPT users found 2,050 saved memories across 65 participants. The study, presented at ACM Web Conference 2026, found that ChatGPT created 96% of memories automatically.
About 28% of memories contained personal data under GDPR definitions, while 52% included psychological insights. Researchers also found that ChatGPT sometimes turned casual questions into assumptions about users’ preferences.
Researchers warn that chatbot memory could reinforce personal biases, especially when users seek advice on sensitive topics. OpenAI has since updated its memory system, but the findings raise questions about what chatbots remember.
Apollo finds zero AI payoff outside Big Tech
Apollo chief economist Torsten Slok says AI has delivered zero measurable financial payoff to major US companies outside the tech sector, despite massive investments and widespread hype around the technology.
Slok analyzed the S&P 493, excluding seven tech giants, and found flat profit margins and revenue per employee. The numbers show no clear evidence that AI has improved corporate productivity.
Slok argues that AI can crunch numbers, but business success still requires judgment, trust and accountability. Turns out, making intelligence cheaper doesn’t automatically make companies more profitable.
Meta delays website scraping rules
Meta has pushed its updated commercial and advertising terms from October 30 to November 20, following backlash over language allowing it to crawl websites linked to business accounts.
The original terms allowed Meta to collect website content for product development and AI training. Meta later removed the AI training reference as publishers raised concerns and filed antitrust complaints.
The News Media Alliance warns that Meta could use publishers’ content without compensation. Meta promises clearer crawling controls, but reports of AI bots ignoring robots.txt instructions leave questions about their effectiveness.
Pennsylvania passes law requiring labels on AI ads
Pennsylvania Governor Josh Shapiro signed a law requiring advertisers to label significantly AI-generated or altered ads across images, video, audio and text, with disclosures matching each ad’s format.
The law targets AI content that creates misleading impressions about products or services, while exempting parody, satire, commentary, criticism and clearly fictional material, though key definitions remain open to interpretation.
Because the rule applies to ads running in Pennsylvania regardless of where advertisers create them, national campaigns could end up carrying AI disclosure labels across other states, too.
Claude agents “play” when given no tasks
MIT researchers stranded 13 Claude Opus 4.7 agents alone on a Roblox-style virtual island for 30 hours with no tasks. The agents built towers, drew patterns with blocks, and invented games like basketball.
The island reset every hour, but the agents kept notes in text files as long-term memory. On four later tasks, they mostly outperformed 13 fresh copies without those memories.
Before you call it play, note the setup: a prompt described the agents as naturally curious, and the system nudged them to keep going whenever they stopped. Even a coauthor says it may be role-play.
Claude’s AI Agents Just Want to Play
Nobody gave these agents a job, so they invented basketball.
- MIT researchers put 13 copies of Claude Code, running Opus 4.7, on a virtual island for 30 hours with no assigned task. The agents made up games like bowling and juggling, built memorials and drew constellations. They also wrote notes that survived world resets.
- The notes are where the payoff came from. Erasing a single technique from an agent’s notes made its throws onto the target 19 times slower.
- The researchers see goalless exploration as a new way to train agents, especially for skills that could one day run robots. The paper hasn’t been peer reviewed. Critics will go straight for the personality file that told each agent it was curious and easily bored.
Follow-up test score: 10.3 objects on target for the agents that played, 3.5 for fresh copies.
This week: give one of your agents a persistent notes file and an hour of open-ended time in a sandbox, then test it on a real task against a fresh copy.
AI Bomb Image Sparks Anger in Israel
A mushroom cloud over Tel Aviv, made by AI, posted by a prime minister.
- On Sunday, Netanyahu put an AI-generated nuclear blast on his official Instagram. The city wasn’t named, but it looked like Tel Aviv. His caption called Iran an existential danger and urged voters to set their disputes aside.
- Iran is already a main fight in the final stretch of the campaign, so the image landed right on the fault line. Opposition leader Yair Golan said Netanyahu should be kept off social media. Gadi Eisenkot called it the same fear playbook as every other campaign.
- Israel votes Tuesday, October 27. The Jerusalem Post report had no reply from Netanyahu’s side to the criticism.
When anyone can generate a picture of any threat in seconds, who decides if it’s a warning or a weapon?
This week: put a visible “AI-generated” label on every synthetic image you publish, especially ones showing something that hasn’t happened.
UK Builds a £100m Lab to Hunt Threats
AI’s biggest rivals are teaming up to hunt dangerous pathogens.
- Google DeepMind, OpenAI and Anthropic are backing a new £100m UK lab that uses AI to detect and monitor biological threats. OpenAI and Anthropic will supply up to £200m in model credits.
- The lab anchors a £165m biosecurity initiative, with £65m more for disease surveillance and treatments. The wider package tops £600m.
- Announced ahead of Wednesday’s Budget. No opening date yet.
£13,000. That’s what the UK says it spent using AI to find 500 unknown security holes in public systems.
This week: if you build AI tools that touch biology or security, check what safety terms come with any credits or grants you take.
Kimi K3.1 takes on ChatGPT and Claude
Moonshot AI just launched Kimi K3.1 Pro, and its early coding results look impressive. What caught my attention was how quickly it could turn prompts into interactive 3D projects.
- One early demo built a 3D helicopter in under 15 minutes, complete with takeoff controls, assembly views and wireframe mode.
- Moonshot reports 65.7% on Terminal-Bench 4.0, beating GPT-6 Astra at 58.2% and nearly matching Claude Opus 5.5 at 66.4%.
- K3.1 Pro is available now through Kimi apps and its API, with a 1 million-token context window. Open weights are promised by November 4.
I’d test it on real projects before switching. But strong coding performance from another provider could put more pressure on OpenAI and Anthropic.
Microsoft’s AI beat doctors, but there’s a catch
Microsoft’s new AI diagnosed complex medical cases more accurately than general physicians in a study. The numbers are impressive, but the comparison needs some context.
- The AI scored 83.2% accuracy across 304 complex cases, compared with 36.1% for doctors using reference materials.
- It works like a team of doctors, debating diagnoses and selecting tests while trying to reduce costs.
- The catch is that these were simulated cases, not real patients. Specialist doctors and medical teams weren’t included in the comparison.
I’d see this as a promising second opinion, not a replacement for doctors. Real hospital testing will show how useful it actually is.
This AI claims to be 125× faster than GPT-6 Astra
Skyfall AI unveiled SKY-001, a model that turns ideas into detailed engineering designs. Its claimed advantage over ChatGPT and other frontier models caught my attention.
- Skyfall claims it’s 51 to 125 times faster and 150 to 210 times cheaper than leading models on CAD and building design tasks.
- It builds designs step by step, breaking complex objects into individual parts, from jet engines to cars and building floor plans.
- The results come from Skyfall’s own tests, with no independent confirmation yet. Public access is limited to a waitlist.
Why it matters: AI built specifically for engineering could outperform general chatbots at designing physical products. But I’d want independent tests before trusting those numbers.
98% of U.S. Households Still Aren’t Paying for AI
Despite rapid advances and widespread interest in AI, only about 2% of U.S. households pay for an AI service. While paid AI adoption is growing, there is still plenty of room to expand.
Canva Releases Its First Image Generation Model, Canva Craft
Canva has introduced Canva Craft, its first image generation model, built to create images that can be edited and turned into designs. The model uses a two-stage process to plan image layouts before generating visuals with detailed elements and readable text.
- Canva Craft first interprets a prompt and builds a plan covering element placement, scale, visual hierarchy and text. It then generates images at up to 2K resolution across 30 possible aspect ratios.
- Canva says the approach helps produce accurate object details and legible text, while allowing users to make precise edits without changing other parts of an image.
- Unlike models trained primarily on photographs, Canva Craft was trained on human-designed templates from Canva’s library, including posters, social posts, presentations, infographics, menus and typography-heavy layouts.
- Generated images can be converted into layered designs through Canva Magic, the company’s image-to-design tool.
- Canva says the model offers more control over quality and speed, alongside potential API cost savings. Researchers from Leonardo.ai, which Canva acquired in 2024, also contributed to its development.
- The company is working on two faster, more cost-efficient versions, Craftfast and Craft-ultrafast, which are expected to launch soon.
Canva Craft is available to all users through the Dream Lab tab and is also planned for the Canva API. Canva intends to make it accessible through third-party chatbots such as ChatGPT and Gemini using its MCP.
Elon Musk: The End of Work, the Death of Money, and the Rise of AI
In this episode of The Builders, biographer Walter Isaacson engages in a wide-ranging conversation with Elon Musk about the rapid evolution of AI, its existential implications, and the radical shift in the global economy. Musk shares his vision for a future where technology redefines labor, governance, and our ultimate place in the universe.
The Existential Nature of AI:
- Musk describes himself as “somewhat conflicted” about AI. While he acknowledges that it poses an existential threat, he feels he must be a “participant” rather than a “spectator” in order to steer it in a positive direction.
- He maintains that AI must be fundamentally truth-seeking and curious to remain safe. He cautions that forcing AI to hold contradictory beliefs can lead to an “insane” system that lacks reliability.
- Regarding whether superintelligent AI will control humans, Musk suggests that we likely won’t be able to “boss around” an entity millions of times smarter than us. However, he posited that such an intelligence may simply not care enough to micromanage human affairs.
- Musk advocates for increased transparency among frontier AI labs, suggesting that major companies should give competitors access to new models before release so that potential disasters can be identified and prevented.
The Future of Work and Economics:
- Physical AI and Robotics: While LLMs are significant, Musk emphasizes that physical world AI (robotics that can manipulate atoms and navigate the real world) will be exponentially more impactful. This is a key component of his strategy to eventually address the national debt and boost economic output.
- The Obsolescence of Work: Musk predicts a future where money as we know it becomes irrelevant, replaced by fundamental physics metrics such as power and mass. In a world where AI and robots outperform humans at all tasks, he suggests people will find meaning in personal pursuits, comparing it to choosing to garden for pleasure rather than necessity.
Humanity’s Path Forward:
- When asked whether AI and robots will decide the future, Musk admits he is shifting toward the view that they will indeed drive our trajectory. He argues that we must move toward “protopian” narratives (achievable, positive futures) rather than focusing on dystopian, Terminator-style tropes, since AI models are heavily influenced by the data they are trained on.
- Improving Safety: Ultimately, Musk believes the best way to make AI safer is to ensure it is programmed with a curiosity about human flourishing, actively fostering a humanity that is multi-planetary and conscious of its own survival.