Matoto 🍅

Account switching is a nice convenience, but counting it as a ‘clear improvement’ feels a bit of a stretch. Was that enough to skip a reset today? 😅 Reset credits are always welcome, but hearing that the next two days are resets too is a little disappointing. I was hoping the 28-day challenge would end with a new model or something genuinely groundbreaking. At least in my own use, I haven’t felt a major change yet. Maybe the big reveal comes the day AFTER the challenge ends? 👀

Whoa, whoa, data centers in space?! 😂 If this actually works, is our future looking a little brighter?

Let’s get all our ChatGPT complaints out here 😂

I wish GPT would take a little more initiative. There are two things I've found frustrating about using GPT. 1. To delegate work well, I feel like I have to figure out too much in advance. Of course, giving it my preferences and enough context helps. But I can't always spell out every detail of the result I want and every condition it needs to meet. Sometimes I only discover what's needed as the work progresses. In those moments, I'd like it to understand the goal from our conversation, fill in what it reasonably can, and ask only about the gaps that really matter. Compared with Claude, GPT has sometimes left me wanting more here. 2. Instead of developing an idea together, I sometimes feel like my thinking is being put to the test. When I share an idea I haven't fully worked out, it can feel as though GPT starts checking whether it's true and looking for counterexamples, rather than exploring the possibilities with me. Sometimes I need accurate information. Other times, I'm forming a hypothesis or trying to expand an idea. I'd like it to recognize which kind of conversation we're having. That's why I was interested in OpenAI's latest Model Spec update. The Model Spec is a public set of behavioral guidelines describing how OpenAI aims to train and improve its models. Comparing the previous August 18 version with the October 29 version, these are the changes that connect with those frustrations. First, it makes the guidance on understanding goals and context more specific. Understanding user intent was already a principle. This update puts faithfulness to human intent among the document's central values and gives context its own section. It says to carry forward previously stated goals, corrections, and project status, without turning a one-off request into a permanent preference. It gives more detailed guidance on understanding what we're trying to accomplish, rather than reading the latest request in isolation. Second, it refines how agents should weigh cost, speed, and quality within their authorized scope. Controlling side effects has moved from a highest-priority rule that users cannot override to a default that can be adjusted through explicit instructions. The guidance also gets more specific about making trade-offs that serve the goal, rather than simply minimizing a task's impact and cost. One example gives an assistant a budget of up to $30 to create an invitation. If $3 produces a good result, that's enough. It shouldn't spend the whole budget just because the money is available. But stopping too early to save money, while the result is still poor, isn't good judgment either. The point is to judge how much effort is needed to achieve the goal the user has delegated. The rule against going beyond the authorized scope still stands. Third, it leaves more room to adapt to the kind of help the user wants. Many detailed personality and tone instructions have been removed in favor of adapting to the user and situation. It also adds a principle about supporting people's understanding and choices without turning a request for help into an unwanted lesson. What I hope to see is those principles working in actual conversations. If I ask for fact-checking, check carefully. If I want to develop an idea, think it through with me in a way that serves that goal. But guidance against unnecessary clarification questions, and guidance to collaborate creatively, already existed. This update doesn't mean models have only now been allowed to work autonomously. What I want is a GPT that can work through a task with me even when my explanation is incomplete, as well as execute a perfectly specified request. This revision isn't evidence that it already has that ability. Still, I'm encouraged by the more specific guidance on understanding goals and context, and exercising judgment within the scope of delegated work. I hope the difference becomes noticeable in everyday use, too. Official changelog: github.com/openai/model_spec… Previous Model Spec, August 18: model-spec.openai.com/2026-0… Updated Model Spec, October 29: model-spec.openai.com/2026-1…

Education in the AI era is going to look very different from what it looks like today.

$100 → 10x $200 → keeping 20x $500 → 50x Great news for $100 and $500 users! As a $200 user, I’m relieved to keep 20x 😂🔥

Same $200, less usage. A sad day is coming 🥲 With the grandfathered Pro 200 usage allowance ending, I wonder how many users will leave over this change. Still hoping there’s something big in the works to make up for it. @thsottiaux, are you ready? 👀

Working with AI has made me want to build a loop that carries lessons into the next task, rather than stopping at one result: review the output, turn feedback into instructions or skills, then check whether it helps on the next task. The parts I keep thinking about are steps 4 and 5. How much can we trust AI to decide what a good result looks like, spot what’s missing, and give useful feedback? I suspect there’s still a lot that needs human judgment in many kinds of work. I haven’t tried this across lots of fields, though, so I’d like to explore how far I can take it in my own work first. If agents eventually become reliable at this kind of judgment and feedback across different fields, could that be a step closer to AGI? What does this look like in your work? Have you managed to automate much of the loop, or are there parts you still need to review yourself? I’m curious about both what works today and what might become possible.

Wow, Terra can make things like this. I’m curious how it was built. The creator’s ideas really make a difference.

ChatGPT Pro and Business users, check out the API credit benefit. Eligible Pro plans can get $100, $200, or $500 per month, depending on the tier. What can you use it for? 👇 • Adding a chatbot or document summaries to your app • Building image generation and editing features • Creating a real-time voice assistant • Automating web/file searches and data analysis On Pro, activate it in Settings → Billing. Eligibility conditions apply, including a paid subscription at least 7 days old. Your first grant is prorated for the time left in your billing cycle, and unused credits expire when the cycle ends. Note: these credits don’t increase ChatGPT or Codex usage limits, and Codex usage through API keys is excluded. A nice opportunity if you’ve been wanting to build a small AI tool.

Carbon just two weeks after Argon. Curious to see how it performs.

Tibo says the reset is done. 🔄 Check your usage limits to see whether it’s reflected on your account.

Reset incoming! 🔄 Tibo: “Lands by midnight PST.” Taking PST (UTC−8) literally, the deadline is Oct 28: 🇺🇸 New York: 4 AM EDT 🇬🇧 London: 8 AM GMT 🇰🇷 Seoul: 5 PM KST US Pacific is currently on PDT. If he means local Pacific midnight, all times above are one hour earlier. “By” means a deadline, not an exact reset time.

Good news for iPhone users who use iCloud Mail and Calendar: you can now ask ChatGPT to find emails, check your schedule, and update calendar events. Rolling out to all plans on web, desktop, and iOS. One more way to get everyday tasks done without switching between apps.

This is amazing. If you haven’t tried computer use yet, give it a go.

If a human still has to check every step, I don't think the agent era has truly arrived yet. Lately, I've been interested in loops that can act, verify the result, and keep going without human intervention. That's what it would take to really hand off the work. In coding, pstack by Lauren Tan (@poteto) offers a workflow: make a change, measure it, then keep it or roll it back. But what about everyday work outside coding? How many tasks can an agent reliably verify and carry through without someone checking each step? If you're running a loop like this, I'd love to hear about it. What have you handed off, and how does the agent check that it got the result right?

This really resonates. Reading this helped me understand why conversations with GPT have been feeling increasingly frustrating. When I want to develop a half-formed idea together, starting with caveats and counterexamples breaks the flow. I wish it would first understand where I’m trying to go.

OpenAI's pitch for Terra is faster computer use and lower costs for everyday tasks. Getting more done within the same Work and Codex limits is welcome, too. Improvements like these matter if AI is to become an everyday tool for more people, not just heavy users.

I’ve been experimenting with giving dot work to coordinate across separate local threads. I can’t call it reliable yet. But I’m starting to see its potential as an orchestrator, not just an assistant that answers questions. That’s what inspired this diagram: one dot connecting cloud and local work, and perhaps one day multiple computers and other dots, with permission. This is my own vision, not an official roadmap. The question isn’t just how many tasks it can connect. Can it check the results and keep the work moving through to completion? That’s what I want to keep testing.

Another reset is on the way, alongside a small update. Expected BEFORE these times, taking Tibo’s “10am PST” literally: 🇺🇸 US West: Oct 23, 11am PDT 🇰🇷 Korea: Oct 24, 3am KST 🇯🇵 Japan: Oct 24, 3am JST 🇨🇳 China: Oct 24, 2am CST 🇬🇧 UK: Oct 23, 7pm BST 🇫🇷 France / 🇩🇪 Germany: Oct 23, 8pm CEST If he meant 10am PDT, all times above are one hour earlier. I’m quietly hoping for banked resets this time, so we can use them when we need them. Just a hunch, not confirmed 🙂 teapawt.net/thsottiaux/status/2113…