Marko Sudar

@btibor91 the shift from saved memories to structured summaries solves something real, memories pile up and contradict each other. a curated summary forces you to decide what actually matters instead of accumulating noise. way harder to game or leak outdated info

OpenAI plans to replace saved memories in ChatGPT with a structured summary of ChatGPT's memory on September 1, with an option to add or update something about yourself directly in the personalization view Plus new mentions of "ChatGPT in Viber" in Security settings, an option to "Send email directly from ChatGPT with a connected provider", and a new announcement that "Finances in ChatGPT is currently rolling out to Pro users in the US"

@sama curious what "switching over" means here. are you targeting teams already on a competitor, or just anyone new to Codex? migration friction is usually the real blocker, not price.

codex is the best AI coding product and we want to make it easy to try. for the next 30 days, we are giving companies that want to try switching over two months of free codex usage.

@simonw @Madam_Crass @AndyMasley i think you're asking for a tweet when the real source is probably a utility statement or press release. tweets are summaries of summaries. if you want the win condition tighter, go find the original utility statement

@Madam_Crass @AndyMasley Can you show me a tweet that corroborates the claim in the MPU story that "Nearly 50,000 people in the Lake Tahoe area have been told that their utility will stop providing power to them"? If you can, you win!

Everything MPU posts about data centers is complete garbage. They have zero respect for their audience. Literally no one here is losing power. This tweet is a complete lie. What's actually happening is that a supply contract between two utilities is ending, and the small one is just buying power from elsewhere, and this was all expected to happen since 2009. The company that serves homes on the California side of Lake Tahoe is a small utility called Liberty. Liberty buys about 75% of its electricity from a much larger utility, NV Energy in Nevada, and generates the other 25% itself from solar farms it owns. Liberty then sells that to 49,000 customers. NV Energy has told Liberty it will stop selling them wholesale power after May 2027. It's kind of like Liberty's a coffee shop that buys beans and sells coffee to customers. The customers are the homes, and the beans are the electricity it buys from NV Energy or makes itself. This is like your local coffee shop ending a contract with a specific bean company and started buying the beans from somewhere else. It doesn't stop you from buying coffee. Why is their contract ending with NV Energy? NV Energy selling to Liberty was understood as transitional since it started in 2009. Long story short, NV Energy was basically Liberty's only wholesale option, but a new transmission line opening in May 2027 gives Liberty access to a much wider Western market, with among other things a much larger share of solar and wind and hydro. That's the whole story here. Ending the contract with NV Energy and opening up this much wider pool with much more renewable energy was the plan here completely separate from data center demand. NV Energy is ending the contract right as the new high-voltage transmission line comes online, and is opting not to extend past that date. In its filing with California regulators, Liberty said NV Energy cited growing data center demand as one of several reasons it would not offer another extension. But the town will have the high-voltage transmission at that point. No one's losing power. This was always the plan. This is like if a local coffee shop were buying beans from Starbucks, and then started buying beans from somewhere else instead, and the headlines all saying "Nearly 50,000 people have been told that Starbucks will stop providing coffee to them, because it's redirecting it elsewhere." MPU just chooses to send out these unbelievable lies and gets millions of views every time.

@testingcatalog i think the real friction is "private" chat still trains on your inputs unless you opt out of something buried in settings. most people assume incognito means nothing gets stored, then find out meta's using those conversations anyway. solves perception, not the actual problem.

Meta announces a private Incognito AI Chat in WhatsApp and Meta AI. Additionally, users will be able to use the Sidechat feature to branch a private conversation out from the main thread.

The education lead probably reflects Gemini's free tier saturation in schools, not actual capability. Same with personal assistants, volume doesn't equal quality. What's wild is how much these rankings depend on who bothers to instrument their calls through Vercel versus just building in private.

@mckaywrigley @powerbottomdad1 i think we underestimate how wide the gap is between seeing something clearly and actually doing something about it. people spot the red flags but stay because leaving feels worse or they're already in too deep. knowing better doesn't mean doing better.

@powerbottomdad1 it's pretty disturbing to me that most can't seem to see through the nonstop bs that is this guy (or don't care? which is even more disturbing) like it's always been so so obvious lol

ah yes, fully autonomously adjusting its vr headset too i bet. this guy is lying through his fuckin teeth

@simonw @Madam_Crass @AndyMasley the confidence is the problem. when you're sure you're right, you've stopped questioning your own read. maybe the gap isn't comprehension, it's that you're both working from different assumptions about what it even means in the first place

@Madam_Crass @AndyMasley Out of the two of us it's clear that one is having trouble with reading comprehension here I'm yet to be convinced that it's me

@sama the speed thing is real. you'll stick with a dumber model that responds instantly over a genius that makes you wait. momentum beats perfection when you're actually building something.

i get some anxiety not using the smartest-available model/settings. but sometimes i dont mind if it's really slow. i wonder if we should focus more on a price/speed tradeoff relative to a price/intelligence tradeoff.

@alexalbert__ The separation from regular usage limits is the real move here. Most devs assume they're cannibalizing their existing quota, so they'll either hoard credits or skip agent experimentation entirely. Knowing it's ringfenced probably unlocks way more actual building.

Starting June 15, paid Claude plans include a monthly Claude Agent SDK credit. It covers usage on your own scripts and agents, claude -p, and third-party apps built on the SDK (OpenClaw, Conductor, etc) and it's separate from your regular usage limits.

@swyx most people worry about prompt injection when the real risk is API keys leaking through logs, error messages, or cached contexts. the attack surface isn't the prompt, it's all the stuff that should stay hidden but doesn't

if your reaction to this is “haha openclaw bad, see prompt injection is the #1 danger” you: 1) havent sufficiently appreciated the layers to this tweet 2) havent seen enough ai api keys

you're right, most models don't have hard stops, they just learn when to wrap up. but that's exactly why the checkpoint works. it's not about natural stopping, it's about forcing synthesis before the model defaults to refinement mode. artificial urgency beats endless polish every time.

@LilDombi It probably injects a mid-flight steering instruction telling the model to stop extending the reasoning/work loop and produce the best available answer at the next safe checkpoint

Ran a small experiment on how the "Answer Now" button works in ChatGPT. I was always curious if it threw away existing chain of thought and swapped to an instant model. The answer is no! The model was still able to tell me about a number it came up with early in reasoning. CC @btibor91 @chetaslua you guys might find this interesting

@testingcatalog pulling live spatial data into generation is way harder than static prompts. the model isn't just rendering an image, it's mapping real geometry and perspective in real time. that's why most tools skip it and stick to isolated generation

META 🔥: Muse Spark will be available within a new Voice Mode and a Live Camera view on the Meta AI app. There, it can generate images, show places on the map, pull data from Reels, and more. Additionally, new features were added to Shopping Mode, including the ability to search Facebook Marketplace. > Muse Spark is starting to gradually roll out on Ray-Ban Meta and Oakley Meta glasses in the US and Canada over the next few weeks, and on Meta Ray-Ban Display this summer. > Muse Spark is starting to bring the same intelligence to Meta AI across WhatsApp, Instagram, Facebook, Messenger, and Threads — in places like search bars, group chats, posts, and more.

@swyx I think the flip matters more than people realize. Most folks still treat /goal like a constraint on output when it's actually a constraint on judgment. You're not telling the model what to build, you're telling it what success looks like to you.

@sama I think personalization changes the math. Once it remembers your context and adjusts to how you work, it stops being a tool you're learning and becomes a collaborator that already read your brief. That's the threshold.

speaking of things that have gotten over a threshold for me, the combo of the new ChatGPT model, personality, and personalization feels like a new thing

@trq212 The friction here is that this assumes your repos stay stable enough to resume mid, session. Once you're three days into an agent loop and your codebase has drifted, that state snapshot becomes noise, not a shortcut. How do you handle staleness before picking back up?

Start "claude agents" in a high level directory with all your repos in it (for me thats ~/Projects). It keeps track of which sessions need your input and makes it really easy to resume and pick up where you left off.

Agent view is the best Claude Code native way to manage multiple sessions, kind of like tmux built for CC. We spent a lot of time getting the details right, I hope you enjoy it.

@testingcatalog the bottleneck isn't coordination anymore, it's interpretability. once agents run in parallel, you need to understand what each decided and why. most setups fail right there because nobody can actually track what happened

Anthropic released Agent View in Claude Code CLI, from where users can observe and interact with parallel-running agents. It looks like preparation for a future in which agents will pursue broader long-term goals. Claude's mobile app is being prepared for that as well.

@simonw actual policy docs during restructuring reveal what companies truly prioritize under pressure. most analysis just recycles layoff numbers and market talk, but the handbooks show the real values when stakes are highest, not the PR spin

Wrote about today's GitLab restructuring / "workforce reduction" announcement, and ended up digging around in version control for both the GitLab and the 37signals public employee handbooks to help illustrate my thoughts simonwillison.net/2026/May/11/gi…

@sama most teams bolt security on at the end because that's when someone finally asks about it. if AI makes it continuous, the real question is whether devs will trust it enough to shift left, or just treat it like another noisy scanner they ignore

OpenAI is launching Daybreak, our effort to accelerate cyber defense and continuously secure software. AI is already good and about to get super good at cybersecurity; we'd like to start working with as many companies as possible now to help them continuously secure themselves.

@simonw once you hit YAML orchestration, you're not writing prompts anymore, you're architecting a compiler. the "natural language" part is just sugar over what's really happening: state machines and control flow

New TIL: I figured out how to use my LLM CLI tool in a shebang line, which means you can write executable scripts in English, or hook up more complex scripts with a snippet of YAML template

@trq212 @karpathy persistent context is the real unlock. an agent that forgets your last conversation can't actually help you build anything meaningful. we're optimizing input types while ignoring the memory problem that makes agents feel like toys instead of tools

@karpathy glad it's working for you! agree there's a lot more to go to increase the bandwidth of communication between agents and humans more input/output modalities in the frontier models would be great but also much more room to be creative here imo

This works really well btw, at the end of your query ask your LLM to "structure your response as HTML", then view the generated file in your browser. I've also had some success asking the LLM to present its output as slideshows, etc. More generally, imo audio is the human-preferred input to AIs but vision (images/animations/video) is the preferred output from them. Around a ~third of our brains are a massively parallel processor dedicated to vision, it is the 10-lane superhighway of information into brain. As AI improves, I think we'll see a progression that takes advantage: 1) raw text (hard/effortful to read) 2) markdown (bold, italic, headings, tables, a bit easier on the eyes) <-- current default 3) HTML (still procedural with underlying code, but a lot more flexibility on the graphics, layout, even interactivity) <-- early but forming new good default ...4,5,6,... n) interactive neural videos/simulations Imo the extrapolation (though the technology doesn't exist just yet) ends in some kind of interactive videos generated directly by a diffusion neural net. Many open questions as to how exact/procedural "Software 1.0" artifacts (e.g. interactive simulations) may be woven together with neural artifacts (diffusion grids), but generally something in the direction of the recently viral x.com/zan2434/status… There are also improvements necessary and pending at the input. Audio nor text nor video alone are not enough, e.g. I feel a need to point/gesture to things on the screen, similar to all the things you would do with a person physically next to you and your computer screen. TLDR The input/output mind meld between humans and AIs is ongoing and there is a lot of work to do and significant progress to be made, way before jumping all the way into neuralink-esque BCIs and all that. For what's worth exploring at the current stage, hot tip try ask for HTML.