🍄 DAILY SHROOMS

Samsung Beats, Market Freaks — and Suddenly Everyone's Nervous About The AI Trade

Here’s the deal: Samsung Electronics flagged a 19-fold jump in quarterly operating profit, beating expectations, and the market still took a baseball bat to the stock. That is the weird part. This was not a “bad earnings” story. It was a “the bar is now on Mars” story. Samsung has been riding the AI data-center boom, with memory demand pushing profits into monster territory. But investors are starting to ask whether AI infrastructure spending can keep accelerating forever — especially if supply bottlenecks ease and chip-price growth starts to cool. This is not just a Samsung problem. It is a vibes problem. The whole “AI needs infinite chips forever” narrative has been holding up a huge chunk of semiconductor valuations. When one of the biggest memory players posts record profit and still gets sold, the market starts checking the floorboards. Plot twist: this comes as Deutsche Bank’s Jim Reid says AI’s real productivity payoff may still take years to fully show up in enterprises. So you have hardware priced for hyperspeed, economists preaching patience, and investors trying to figure out whether the AI trade is ahead of itself or just taking a breather. The bottom line: the AI trade is not broken. But today was a reminder that it is not bulletproof either.

• POW!DeepSeek is reportedly building its own inference chip, according to Reuters — because relying on Nvidia (and US export rules) is apparently getting old fast.
• ZOOM!The Bank of England warned today that growing AI use in financial markets is becoming a real systemic risk regulators need to watch closely.
• BAM!AI giants are reportedly handing out massive free cloud compute credits to startups right now, basically fighting a loyalty war before these companies even pick a side.

Heyalo

An in-call AI sales assistant that listens during your meeting, but — key detail — mostly stays quiet. Instead of joining as a weird bot guest or dumping notes after the call, Heyalo surfaces buying signals, objection counters, and sourced answers in a private panel while the conversation is still happening. Afterward, it writes the recap, deal signals, and follow-up draft. The hard engineering problem here is not making the AI talk. It is teaching it when to shut up.

ASK in the Dark: Uncertainty-Gated LLM Assistance under Partial Observability

This paper is basically about teaching reinforcement-learning agents when to ask a language model for help — especially when the agent cannot see the full state of the world. The researchers found that vanilla “ask the model when uncertain” methods can fail because the language model does not get enough context. Their improved method, ASK+, gives the small language model trajectory-aware context like a partially revealed map, visited positions, and action history. Why it matters: this is the kind of boring-sounding research that becomes important if we want AI agents and robots to operate in messy real-world environments without asking for help every five seconds — or confidently doing something dumb.

Read the paper → · via arXiv AI Papers

Ford

Here’s a plot twist for the “AI replaces workers” crowd: Ford reportedly brought back hundreds of veteran engineers after AI tools could not fully replace deep manufacturing judgment. The details are nuanced — Ford is not abandoning AI. It is using experienced engineers to make the AI systems better, mentor younger teams, and bring back the institutional knowledge that spreadsheets forgot to price in. Call it a success story for human expertise, and a cautionary tale for anyone who thinks AI is a plug-and-play replacement for people who have seen the weird edge cases in the real world.

No new flagship model drops today — but the spending war says everything

No headline model release in the last 24 hours, but read between the lines: AI labs and cloud providers are handing startups massive compute credits to lock them into their ecosystems before they pick a stack. That is the real API story right now. It is not just who has the best benchmark score. It is who gets embedded into your product, your infra, your workflows, and your burn rate first. If you are building this week, take the free credits. But architect like a mercenary: clean provider abstractions, swappable models, portable evals, and no hard dependency on one vendor’s magic endpoint. Vendor lock-in dressed up as generosity is still lock-in.

Chip jitters are the real story, not just Samsung's earnings

Samsung’s selloff this morning triggered a broader AI-chip gut check, and that matters more than the headline number. The entire AI trade is basically a bet that chip demand stays enormous for years. When one of the biggest memory players posts monster profit and still gets punished, it forces investors to ask whether expectations have outrun reality. Layer on DeepSeek building its own inference chip, the Bank of England warning about AI-related financial-stability risks, and Deutsche Bank saying productivity payoffs may still take years, and you get a market recalibrating expectations in real time. Smart money is not fleeing AI. It is getting pickier.