There's a split from neuroscience I can't stop thinking about as the real reason AI agents fail inside companies. Treat it as an analogy, not a literal claim, but it keeps holding. Your brain runs two memory systems (Complementary Learning Systems theory, McClelland et al. 1995). The neocortex learns slowly and holds general, world knowledge. The hippocampus learns fast: it captures specific episodes as they happen, then consolidates the ones that recur into durable, reusable procedure. A pr...
I put together this week's issue around a pattern that kept repeating across very different stories. Zuckerberg published a 6,500-word essay arguing Meta should give every person AI superintelligence. Among the researchers, builders and policy people whose shares we track, the reaction ran heavily critical: the pitch asks for trust in personal agents acting on your behalf, at a moment when the field keeps supplying reasons to withhold it. The same week: Anthropic's second company-wide risk r...
Most of the reasoning gains coming out of the big labs are still tied to scale. More params, more compute, better reasoning. That's been the play for a while. Ran into TwIL-LM2 which flips the script for narrow tasks. PEFT LoRA adapter on SmolLM2-1.7B, specialized purely for formal logic translation. On strict-7 scoring (no partial credit, exact-format required) it hits 0.2386 - ahead of Qwen3-8B at 0.2093 and Gemma-4-26B at 0.2050. On the loose-match six-lane average it's a different story (Q...
Headline says "bans humanoid robots, targeting China." Neither half of that is quite right. It's not a ban. It's an addition to the FCC's Covered List, which blocks new models from getting FCC equipment authorization. Anything you already own keeps working. The government's exempt too. And it doesn't name China. The FCC's own wording is "place of production, not by entity". A humanoid built in Vietnam gets caught by the same rule as one built in Shenzhen. China's obviou...
So basically modern AI using classic AI is like a human using a calculator?
Comments URL: https://news.ycombinator.com/item?id=49322324
Points: 1
# Comments: 0
So much to read, even for a simple fix or feature. Things like "gate X behind Y", "surface", "escape hatch" and many complexe english wordings that make one mad. I've head a bout an effective proxy py solution but not remember the name of it
Comments URL: https://news.ycombinator.com/item?id=49322520
Points: 2
# Comments: 0
Bloomberg's own reporting already answers the question this clip raises — is the "messy middle" projected or already happening? At Commonwealth Bank of Australia, Microsoft, Uber, and Hyatt, it's already happened: sizable call-center headcount cut using automated phone and chat systems, savings already banked. Economist Molly Kinder's point isn't a forecast. It's a line item that's already closed. I've watched this exact math play out before superior technology ever touched...
Chart uses Ramp AI Index data, discussed by a16z. Spend includes LLM subscriptions, coding agents, API usage and GPU cloud spend. The top 1% line is wild but the median is almost more interesting. Looks like most companies are still experimenting while a small group have turned AI into a serious operating expense