And what a time to enter AbcCVX right now. U get to take advantage of the discount on ur way in too, currently its a bit wider than been for awhile. Result? Even more CVX working for yah, autocompounding and up to 40% APR.. (veCLEV Max boosted) Higher than vlCVX, as usual.. This is my vault, this is my yield, this is my boost!

Upload is a documentary series, not fictional.

Wen live betting as a service? Agentic hamster racing!

I have a weekly Codex quota. Just let me fucking use it. Why is there a 5 hour babysitter sitting on top of it? Yesterday I still had more than 50% of my weekly quota left. Didn't matter. The 5h window was the actual bottleneck. And I had a full reset expiring, so I ended up using the reset before it disappeared. That sounds like free compute. It isn't always. A full Codex reset also resets the weekly window and moves your next weekly reset date. So if I still had quota left that I was planning to burn later in the week, that remaining capacity does not stack. New week starts. Old schedule is gone. I have literally had these “gifts” reduce how much Codex I could use over the period compared to if OpenAI had just left my quota alone. Around 15-30% in some cases from my own tracking. Because I don't use compute evenly. Of course I don't. Some days I barely touch Codex. Then I hit a build, research or debugging run where I want to absolutely cook it for 10 hours. That is what a weekly quota is perfect for. Give me X compute for the week. I decide when the fuck I need it. Instead Plus gets a weekly budget with a smaller timer sitting inside it telling me how quickly I'm allowed to spend my own allowance. And apparently this isn't some fundamental safety requirement. Pro 100, 200 and 500 don't have the 5 hour Codex limit. So what exactly are we doing here? Is the assumption that a Plus user can't be trusted to manage a weekly quota? Because I can. I am actively managing it already. Yesterday I was trying to burn compute I did not need yet because a reset was about to expire. Today I want to burn hard and the 5h limit matters more than the weekly quota again. This creates completely backwards incentives. Save quota for a hard workday? You may not be allowed to spend it fast enough. Get a promotional reset? Better calculate whether using your “free” reset destroys quota you were already planning to use. Get a global reset at the wrong point in your cycle? Congrats, your reset calendar moved. I don't want more babysitting tools for this. I want one number. You have this much Codex compute left. Reset on this date. Go nuts. When it is gone, it is gone. That I understand perfectly fine. If Pro users can manage their compute without a 5h leash, let Plus users do the same. I paid for a bucket. Stop telling me how fast I am allowed to drink from it. @thsottiaux

I want to tell AI what I want done. I do not want to learn why request #83 got a 429. Yet somehow we are normalizing this. 429 means one thing. 403 means something else. 504 might mean the model did the work and the transport died afterwards. A reply can stop because it ran out of output tokens. Another can hit an iteration ceiling with the final answer already sitting there. Same experience for the user. AI broke. We went through 17 terminal failures inside YieldBrain recently. Not one was the model failing to think. 5 were tool or harness failures, 4 rate limits or quota, 4 transport, 2 output budget, 1 auth and 1 insufficient evidence. Zero semantic model failures. That is insane to me. Not because plumbing fails. Of course it fails. Because we still expect the user to know what kind of plumbing failed. Take a 429. One NIM model can be unavailable while another NIM route is fine. We built the fanout so one model hitting 429 does not block the others. If I see that failure, I know not to declare NIM dead. Why the fuck should the user know that? Same wif retries. Some failures should back off. Some should reroute. A retired model route should not be retried at all. We even found an 83 character harness error in the database that could be counted as an answer. Imagine asking AI to do something and the system quietly records “yep, replied” because an error stub exists in the reply field. kek. This weekend made the problem painfully obvious in our own system. On Friday I said publicly that we would release the first open source part of YieldBrain this weekend. The kernel. We did not. It is still not publicly released. The Director got the goal on Friday and kept working on it through the weekend. The brief told it not to bulldoze through failures just to satisfy the release goal. If something useful was repairing itself, leave it running. If a provider looked dead, verify it. If a harness broke, repair that lane first. And public actions still belonged to me. So the release kept uncovering work. Fix one thing, verify it, then find the next thing behind it. That does not fit neatly inside “this weekend.” I was the one who said this weekend anyway. The machine had a goal. I made the promise. By Sunday the kernel source had been promoted into our canonical internal control, but public release was still disabled. There were still decisions around licensing, the namespace and the sanitized public export. Those are human decisions. So it stopped there. Which is actually what I asked the system to do. This is not the part where I claim our AI fixed itself and everything worked perfectly. It didn't. YieldBrain still works too often because I know what the fuck I am looking at. I know that a 429 on one route is not automatically a dead provider. I know why a 410 after that changes the diagnosis completely. I know an output limit is not a network failure. I know when the thing claiming to be an answer is actually garbage from the harness. Of course I know. I built the thing. The user didn't. That is the bug. I don't want to build a better cockpit and teach everyone to become a pilot. Fuck that. The user should be able to say what they want done. Then the system should deal with as much of the ugly shit underneath as it can. If one route dies, it should know what that means before bothering the user. If work can continue somewhere else, continue. If the evidence is garbage, don't pretend the job succeeded. And when it reaches something that actually needs the human, stop. That is much closer to what I mean by “just works.” A hosted coding subscription can own more of this plumbing because it controls more of the stack. YieldBrain is different. It coordinates models, tools and compute the user already legitimately has access to. So right now we see more seams. I want fewer of them. Not by pretending the machinery isn't there. By making the system understand enough of its own machinery that the user doesn't have to. That is also why putting the kernel out publicly matters to me. As long as this thing only works in my setup, with me standing next to it knowing where all the weird bodies are buried, we have not proven the thing I actually care about. Can the knowledge leave my head? Can someone express a goal without first learning our little religion of queues, routes, retries and provider weirdness? Not yet. That is the honest answer. There are roughly three hours left in the weekend while I write this. Maybe the kernel goes public tonight. Maybe it doesn't. I am not making another promise. But the attempt already found the thing I wanted the release to test. Too much of YieldBrain still works because its architect knows how YieldBrain works. The user has better shit to do.

Ohh my… The nerfing of the chat mode in ChatGPT is real… I used to be able to work for days, even weeks in the same chat. Now it reaches Max lenght in a day. Are we compute restrained much @thsottiaux /@sama ?? Wtf is going on?

I have always had this weird belief that if I can picture where something logically ends up, reality will eventually get there. Timing? I suck at that part. But I remember when mobile phones started becoming normal and thinking… this thing is obviously going to eat everything. Camera, maps, music, internet, payments, whatever. Why would we keep carrying separate shit around forever? I feel the same way about AI now. People are learning prompting, agents, memory, routing, context, model selection. I think most of that disappears. Not because it isn’t useful now. It is. I just don’t think the user was ever supposed to know about it. Same thing wif DeFi. Normal people are not going to learn bridges, gauges, wrappers and 14 different ways to move yield around. Why would they? I want more money because I want more control over my own time, so I can spend it on what I actually care about. Don’t fuck me on risk. GO. Everything after that should eventually become somebody else’s problem. Probably a machine’s. I have been wrong on timing before and I’ll be wrong again. But when I can see a version of something with way less friction, I tend to assume we’re eventually going there. If I can envision it, I expect reality to catch up.

Hmm never seen so much errors with chatGPT chat mode… Streaming errors all around today. Hunch: they are replacing the chat mode, with DOTs… Your always on mfr, your one stop connection to all things gpt. @thsottiaux ?

Update on this. The useful part is getting clearer. You shouldn’t need to choose models, manage agents or babysit retries. You tell the system what you want, why it matters, and what outcome you need. The part I’m extracting is the machinery that turns that into work, routes it, checks what actually happened, and decides whether the goal is done. That’s the part I want to make portable. Not calling it shipped yet. Still testing the clean path from goal to verified result.

Google is baked, 4.20.

The computer as we know it will not come back. It will be a small gadget we talk to, which will then render/build it all live there and then. Majority of compute, will still be centraliced, and not with the average user. The feew who builds, serve the rest.

This will be the future, no matter how u look at it. Frontier models will be the cost, the others will be free, or close to free. Model routing, job to fit lower models is the gwei. Already building my system like this, using free providers like kimi k3, glm 5.3 deepseek 4.1 and others for free, while my ChatGPT web chat is the main orchestrator managing Hermes fleet and other harness in the system. On a 20$ gpt plan, i cant spend it all atm, there is just to much compute available.

Westworld is a documentary show, not sci-fi.

Whoa, Niice! Yieldbrain Director Pushed it further. Controll it all from ChatGPT app, using the chat mode which is free. Dm your agents, directly from ChatGPT app, on mobile.. Dispatch missions, automate stuff in the ChatGPT eco… The quota extender of free compute.. My system showed a savings of 2k$ last couple of months using my free compute system..

The professor that really solved it has not acepted the bribe to be quiet yet. Like last time?

Would u want to run such systems for free anon? Cause thats what Yieldbrain Director is doing. A Multi harnes/model/provider/tool/persona control plane using free acsess points as main driver. Its model/provider agnostic, so u can run the control plane from your paid subscription if u wanna, i use my gpt sub, but i use the free chat mode that do not use any quota.. Thats GPT 5.6 SOL on high reasoning as the Director. Its the 20$ sub so thats why its only on high reasoning and not the pro version. Hermes profiles run on free glm 5.3 entrypoints using MAX reasoning. With subagents/graphs/loops built in aswell. All free, and premium models. Kimi k3, deepseek 4.1 flash, glm 5.3 flash.. Wen i want/need more SPEED, i draw from my gpt limits using the real top frontier, gpt 6 Astra xhigh. And no, i do not struggle with quota limits like others, my system doesnt use enough of my quota yet so we train now on finding the right uses for this to Max that value. There is bugs, and far from perfection but it self improves every day. The control plane transformator my intents/wants/needs to downstream Jobs i do NOT need to understand… Wanna build WiF me?

This is exactly why I’ve been watching f(x) and FX100 from a different angle than most perp discussions. Everyone compares fees, leverage, liquidity and UX. I care about what happens when someone decides you shouldn’t be allowed through the front door anymore. With @protocol_fx , the normal website isn’t the protocol. There are decentralized frontends. There is an SDK. And underneath all of it are public smart contracts you can interact with directly. Kill one UI and the position machinery is still there. You don’t even need a UI if you know how to talk to the contracts. That’s a very different architecture from building an onchain exchange where one dominant frontend still decides who gets access. This Hyperliquid screenshot is basically the perfect example. The wallet wasn’t magically deleted. The blockchain didn’t disappear. A third-party risk system told the frontend not to let that address through. Fine. Now ask the better question.. If the main frontend says no, can I still use the protocol? That should probably be part of the decentralization test for every DeFi venue. f(x) already has the pieces I want to see decentralized frontend access, open contracts, SDK access, direct contract interaction. @FX100Perp can carry that philosophy into a much bigger trading surface. The next generation of onchain trading shouldn’t just self-custody your funds. It should make the website optional. Because if somebody can switch off your access by switching off one door, we didn’t decentralize the venue. We decentralized the backend.