Godfred S.

“AGI has arrived.” I’m not convinced. Forget ARC-AGI, FrontierMath, coding benchmarks, and how many GPUs were used. Give the AI a much simpler test. Walk into an ordinary house. Clean it properly. Decide what is trash and what should be kept. Find the old TVs, computer monitors, video cameras, hard drives, cables, and other unwanted electronics. Figure out which can be donated, recycled, or discarded. Find the appropriate local recycling facility, transport everything there, and finish the job without requiring a human to supervise every step. A human housekeeper or junk-removal worker can understand that assignment immediately and deal with hundreds of unexpected situations along the way. If an “AGI” cannot reliably accomplish such an ordinary real-world task, what exactly does 99.9% on a benchmark prove? 100,000 GPUs—or 500,000 GPUs—may produce extraordinary computational capability. But more compute is not automatically more general intelligence. I’ll be more persuaded by AGI when it can handle an ordinary Saturday household job from beginning to end. The ultimate AGI benchmark is not another benchmark. It is the real world. #AI #AGI #ArtificialIntelligence #GenerativeAI #FutureOfAI #Robotics #Automation #RealWorldAI #AIInnovation #FutureOfWork #LLM #MachineLearning #NVIDIA #OpenAI

This has to be the best explanation of vibe coding 👏😂 → Random rails going nowhere - just like code Cursor generates → Detours, loops, backtracking - just like Loveable → Goes off the rails… then somehow fixes itself - just like Claude Code → Technically works. Forever. AI doesn’t write code the way engineers do. It writes code the way a raccoon builds a nest: → Whatever fits → Whatever sticks → Whatever works right now Is it clean? No. Is it elegant? Absolutely not. But does it run? Yeah 😎 --- Battery-powered sensors have to sample sparingly. That’s the whole game: save energy while still reconstructing the signal between measurements. And the decision logic has to fit on microcontroller-class hardware. So we used a large language model as an offline policy compiler. You give it a natural-language task description and a compact rule grammar. It hands back policies as a few numeric thresholds and integer periods, under 100 bytes. No weights are fitted. The model never sees the raw series, only eight summary statistics. Training-free but statistics-informed. The policies come from an agentic loop: a generator LLM proposes candidates, a deterministic evaluator scores each one exactly on the recorded trace, and an optimizer LLM refines them over eight rounds. We evaluated Claude Opus 4.8, Claude Fable 5, ChatGPT 5.4 and ChatGPT 5.5, five seeds each, on the Beijing PM2.5 series. Against classical baselines, a budget-matched random search, and NSGA-II and ParEGO over the identical grammar. The loop improves on zero-shot generation and random search. Pooled Wilcoxon p<10-5. Within a 0.20 duty-cycle budget, the best model’s mean across five seeds hits an RMSE of 21.79 versus 25.09 for random search. And 15.98 versus 20.57 within 0.30. The gain narrows at the tightest subsampling. An interpolation-error bound explains why, through the signal’s curvature. https://lnkd.in/ePfbwjsC

What caught my eye here isn’t that tabular models are new. This has been building in research for a while. The shift is that Fundamental is trying to ship it as an enterprise product at scale, with enough capital and distribution to prove it out. The real test is simple: does it beat well tuned XGBoost and LightGBM across messy, uneven real world tables, and can it run inside governed data environments where latency, cost, and explainability matter. I don’t think this replaces what teams use today. The upside is pairing a tabular model for prediction with a language model for context and workflow. If that combo lands, it could matter more than the headline hype.