Yuma Kakuya | AI Agent Systems @null_founder

@Kamil_Sierocki I look back on the day, then imagine myself succeeding tomorrow. It sounds simple, but ending the day with that picture in my head helps me reset and start again.

@headzoo I know this feeling 😂 Once enough agents are running in parallel, you eventually discover work you forgot you delegated. It’s a great surprise, but also exactly why I care so much about orchestration, task ownership, progress logs, and verification.

This is exactly why I’d pick distribution if I had to choose one moat. Frontier models can change every few months. But being embedded in the workflow, identity, cloud, productivity stack, and daily habits of millions of users is much harder to replace. Microsoft doesn’t need to win every model benchmark if it owns the route to the user.

Yes, that’s exactly the failure mode I’d treat as a hard harness invariant. With Fable 5.1, the thinking block is prefix-bound. If keep-tail compaction rewrites the top-level system prompt, tool definitions, or any earlier message, replay can 400. So the safe design is append-only: preserve the prefix and thinking blocks exactly, compact around them rather than rewriting them, and regression-test replay after every compaction path. If the prefix must change, I’d treat that as a new execution branch rather than pretending it’s the same continuation.

🚨 THE FRONTIER JUST SPLIT INTO SPECIALIZED LAYERS. ONE MODEL DOESN’T WIN EVERYTHING ANYMORE. Claude Fable 5.1 just opened a massive gap in ultra-long-horizon coding. OpenAI just turned voice into a separate realtime layer that can delegate the hard thinking elsewhere. Those two updates point in the same direction: THE FUTURE IS NOT ONE SUPER-MODEL. IT’S AN AGENT STACK. Two developments matter. 1/ FRONTIERSWE V2 JUST EXPOSED A HUGE LONG-HORIZON CODING GAP Proximal expanded FrontierSWE from 17 to 34 tasks. Each model gets: → 5 trials → up to 20 HOURS per task → compiler optimization → scientific computing → AI research → visual reasoning → long-running implementation and repair This is not “generate the function correctly on the first try.” This benchmark rewards the ability to: → plan for hours → recover from failures → preserve state → interpret feedback → decide when to checkpoint → keep making useful progress And the gap is brutal. FrontierSWE v2: Claude Fable 5.1: 56.3% GPT-5.6 Sol: 32.2% GLM-5.3: 30.2% Kimi K3: 25.9% Grok 4.6: 25.3% Gemini 3.7 Flash: 20.3% Qwen3.8-Max: 15.8% DeepSeek V4 Flash Vision Exp: 14.8% Fable is more than 24 points ahead of Sol. That matters because short coding benchmarks increasingly make frontier models look close. Give them 20 hours and the separation comes back. The real frontier is becoming: NOT “who writes code best?” but: “who can stay coherent, recover, verify, and keep moving for hours?” There is one important caveat. Fable 5.1 falls back to Opus 5 on tasks blocked by content filters. So 56.3% should NOT be treated as a perfectly pure Fable-only score. But even with that caveat, the long-horizon signal is hard to ignore. And GLM-5.3 may be the other surprise. 30.2%. That puts an open-weight model extremely close to Sol at 32.2%. But efficiency still separates them. According to Proximal: Sol averages about 8.6 hours and 182M tokens per trial. GLM-5.3 uses roughly 2x the time and 83% more tokens than Sol. Still, among models above 30%, GLM-5.3 was the cheapest at roughly $97 per trial. That gives me a very different routing picture: Hardest long-horizon parent / Senior: → Fable 5.1 Fast coding / integration: → Sol Large-scale open-weight workers: → GLM-5.3 But the model itself is only half the story. Proximal also found that the SAME model can improve when run through a better long-horizon harness. Compaction. Progress logs. Resubmission. Remaining-time awareness. Verification loops. Those are no longer orchestration details. They are part of the benchmarked capability. 2/ OPENAI JUST SEPARATED REALTIME VOICE FROM FRONTIER REASONING GPT-Live-1 is now available through the API. And this is much more important than “better voice.” GPT-Live-1 is a full-duplex realtime model. It can listen while speaking. So instead of: Speech → STT → LLM → tools → TTS you can now build: GPT-Live-1 = realtime interaction layer ↓ Astra / Codex / Luna / third-party backend = reasoning + tools + execution That means your voice interface no longer needs to wait for your smartest model to finish thinking before the conversation can stay alive. OpenAI is explicitly separating: LATENCY-SENSITIVE INTERACTION from HIGH-INTELLIGENCE EXECUTION. GPT-Live-1 can become the traffic controller. The backend can be whatever is best for the task. Coding? Send it to Claude Code or Codex. Cheap retrieval? Use a cheaper worker. Long context? Route elsewhere. Deep reasoning? Escalate to Astra. This is exactly where multi-model orchestration gets interesting. OpenAI reports GPT-Live-1 beating GPT-Realtime-2.1 by +30 percentage points on Full Duplex Bench. And GPT-Live-1 + Astra Medium reportedly reaches #1 on Tau3. Those performance claims are still vendor-run, not independently replicated. But the architectural signal is already real. Put both updates together: FrontierSWE says long-running engineering amplifies model differences. GPT-Live-1 says the realtime interface doesn’t need to be the model doing the deepest reasoning. So the stack starts looking like this: Realtime interface ↓ Router ↓ Long-horizon parent ↓ Specialized workers ↓ Independent verification ↓ Execution authority The frontier is no longer converging toward one model doing everything. It is converging toward specialized intelligence assembled into one Agent system. And that is a much more interesting race.

@donidhernandez This is the part I’m paying more attention to now. Attention and followers are useful, but they don’t automatically become customers. The real work is building enough trust, qualification, and follow-up between interest and purchase.

@Wearcoby I relate to this a lot. I don’t think persistence means blindly doing the same thing forever. It means staying in the game long enough to keep learning, adapting, and improving while everyone else slowly drops out. That part matters more than people think.

@ChrisTranGG “Overqualified” often sounds like a polite way of saying: “We’re worried you’ll see the problems too quickly, want more than we can pay, or leave once you realize how this team actually works.” Sometimes the rejection is the warning.

@fullstacklover Probably a perfect abstraction for code that gets deleted tomorrow 😂

@ayoub_el77927 I’d spend more time actually being present on X and talking to people who seem to share similar goals. I’m still far from your numbers, but I’ve finally realized how much “every day” matters. Show up, talk, learn, repeat.

@shakhzod1001 My actual life is pretty morning-oriented. My X activity, though? Definitely a night owl. It’s 5 AM here in Japan right now and I’m still on the timeline 😅

Sources / evidence: FrontierSWE v2 — official announcement & methodology frontierswe.com/blog/v2 FrontierSWE v2 — live leaderboard frontierswe.com FrontierSWE — changelog frontierswe.com/changelog OpenAI — GPT-Live-1 API launch openai.com/index/introduc… FrontierSWE is an independent benchmark operated by Proximal. Important caveat: Fable 5.1 uses Opus 5 fallback on tasks blocked by content filters. GPT-Live-1 performance claims on Full Duplex Bench and Tau3 are currently OpenAI-run, not independently replicated.

@borrhensaidi There are plenty of days like that. I’m still nowhere near your numbers, but I believe there’s a day when the work gets recognized if you keep showing up. Connected with you. I’ll be keeping an eye on your timeline from here on out.

@thsottiaux @thsottiaux I barely used Spark when it first arrived. But somehow, once you hear it’s being retired, it gets a little sentimental. Funny how even model names start to feel like part of an era.

@ChrisTranGG Distribution is probably my biggest weak point right now. Building is the easy part for me. Getting the right product in front of the right people consistently is much harder. That’s the task I keep having to force myself to do.

@VicNexAI I’m pretty confident too, but it’s still a prediction for now. If GPT-6 Sol actually drops today, even after all these signals, I’m still going to be shocked 😁

🚨 I THINK GPT-6 SOL COULD DROP TODAY. Sources are in the self-reply below.👇️ Not “sometime soon.” TODAY. We still don’t have official confirmation from OpenAI, so this is a prediction. But the signals just got much stronger. A new model listing has surfaced showing: gpt-6-sol Platform: OpenAI That matters because this is no longer just: “Someone says OpenAI is testing a new Sol.” We now have a production-looking model ID appearing on an actual model surface. And the timing is interesting. Right now, OpenAI’s public API model catalog still shows: → GPT-6 Astra → GPT-5.6 Sol → GPT-5.6 Terra → GPT-5.6 Luna There is still NO public gpt-6-sol. The API changelog also has no GPT-6 Sol release entry yet. So we appear to be sitting in the gap between: model appearing somewhere in the infrastructure and model becoming officially public. That gap can be very short. There’s another signal worth watching. On September 3, the same day GPT-6 Astra appeared in OpenAI’s API changelog, OpenAI’s Codex repository received a commit explicitly adding a hidden: gpt-6-astra model definition to its bundled model catalog. That commit added reasoning levels, tool capabilities, context limits, Agent instructions and review policies. Today, I checked the public Codex repository for: gpt-6-sol Nothing yet. That makes the next Codex commit extremely interesting. If gpt-6-sol suddenly appears in: → OpenAI API model catalog → OpenAI API changelog → Codex bundled models → official OpenAI developer channels then the switch has effectively flipped. My current guess: GPT-6 Sol is already somewhere in the release pipeline, and there is now a real possibility OpenAI releases it during the U.S. day on September 11. For Japan, that means: late tonight → September 12 morning. Could it slip? Absolutely. This could still be staging, an internal catalog, limited exposure, or something that sits there for days. But if I had to bet right now: I would bet on GPT-6 Sol arriving much sooner than most people expected. Possibly today. I’m watching the release surfaces.

私も、タイムラインの数字が2桁のままで悲しくなる日がありますし、大きくバズったことも、突然Spikeしたこともありません。 だから、とにかく毎日続けています。 毎日発信して、他のbuilderが何を考えているのかを読み、いろいろな意見を聞いて、そこから学んだことを自分の仕事に活かせるようにしています。 たぶん、今の自分が一番信じているのは、その継続です。

@divyakumar_p 今日は、やっと順調に動いていると思っていたアプリが、実はまだ完全にポンコツで、画面に八つ当たりしそうになりました 😂 本当に毎日が勉強です。

@MaberFate You just described a dangerous percentage of indie builders on X 😂 I’m building AI agent infrastructure and small developer tools under StereoBinary. Still very much in the “turn building into revenue” phase, so yes, we should probably connect.

Hey 👋 I’m building Void. My Mac died, so I’m shipping Windows and Linux builds for now while I bootstrap my way toward replacing it. void.stereobinary.com Would love to connect with people in tech who are building through setbacks and trying to fund the next step themselves.