We have AGI-capable models today already. You just need to innovate the proper harness infrastructure. The top 10 AI users on the planet haven’t even cracked 1% of intelligence-extraction for AI. What do you think happens when someone hits 10% in the next x months? Memory, looping, signal-extraction, objective continuity, non-colliding parallel agent orchestration, etc.. There’s so many layers that nobody has even cracked 1% of the “peak actualization state” of. If someone even cracked 10%, people would freak out and say all knowledge workers are getting replaced effective immediately, today. Better models just make the mean/average response better, which normies can extract that nice little 0.01% intelligence sliver of without any sophisticated harness. You’re all applauding capability increases of that unskilled 0.01% intelligence extraction. If they stopped releasing models forever, today’s SOTA GPT-5.4 and Opus 4.6 are enough to replace virtually all entry level knowledge workers. Nobody is even 100x close to proper harness/orchestration yet.

6.1 Sol Ultrafast: 283 TPS 6.1 Sol Default: 40 TPS 5.3x - 7.1x faster. Not bad. Better than a couple weeks ago; Ultrafast used to yield only ~200 TPS. Still not the 750 TPS we were falsely promised 6-8 weeks before Ultrafast finally launched.

240B tokens with Opus 5.5 within the last 18 days. Not bad. (Yes, I have several $200 Claude accounts)

SlopenAI

Sol 6.1 High just ran for 36h. Nonstop computer use, using iPhone mirroring, for my iPhone 16 Pro Max. It checked every single hardware configuration option possible for: - MacBook Pro - MacBook Air - MacBook Neo - iMac - Mac mini - Mac Studio Looked through all possible builds (that wouldn't take 2-6+ weeks to ship/pickup). It found a 48 GB RAM MBP, and 64 GB RAM Mac Mini. All available for immediate pickup. Few other finalists in second image. (I'm optimizing for RAM). 36h task but used around 15% usage. Insane. Sol-6.1 High is killer

It’s wild how I created a version of RepoPrompt.. 6+ months before it even existed.. With the code being 100% authored by AI. By the end of o1, it was well over >25K LOC But I just didn’t publicize or distribute it I wonder what else I could’ve done differently and publicized (considering how, I’ve created multiple other apps adjacent to it, for an entire dev workflow—SOTA at the time) Plus dozens and dozens of other apps and tools I could’ve easily open sourced or something Sigh