AI & Tech Daily News, October 12, 2026
This day's round-up covers five stories that run from the disconnect between hype and reality in language models to practical tricks for cutting token costs. There is also a look at mathematical proofs, a sharp critique of AI IPO valuations, and a surprising WebAssembly port of PlayStation 2 games. Each source link is below.
First on my mind is The gap in shared understanding of LLM capability is widening. Andrew Karpathy has a way of naming things, and this one lands. The mismatch between how people picture what language models can do and what they actually do is not a minor confusion. It shows up in how teams set deadlines, in how reviewers judge output, and in where the public draws the line between tool and mind. The news cycle feeds both extremes: treat a model as either a genius or a parrot depending on which day you read it. Neither picture is right, and the wrong picture costs real time.
Then there is the everyday work of making agents affordable. Ted Haley writes about local indexing to reduce tool calling and token usage, and it offers a reminder that the most expensive part of an agent is often the chatter. Every tool call sends more tokens through the wire and waits on a response. The trick of preloading the facts an agent is likely to need into its own window changes the economics instead of patching the prompt. You get the same answers for less of the burn. That is the sort of quiet win that compounds across hundreds of runs.
On the theory side, Terence Tao: Math 2.0 asks what happens when proof work is shared with models in the loop. Tao is one of the few people whose name opens a door in every subfield, and this presentation does not promise a replacement for rigor. It sketches a setup where computation and human direction meet without handwaving. The slides are a starting point, not a finished system, but the question is the interesting part: how much of the creative labor in math can actually be offloaded and how much must stay in the human head.
On the finance side, a research firm is calling Anthropic's IPO the most ridiculous of the year. The analyst's point is that the projections baked into the pricing rest on enthusiasm that may not survive market pressure. That is a normal argument, but it comes hard on the heels of so many AI valuations that the repetition matters. Prices today are betting on a demand curve most people have not seen yet. That is not obviously foolish, and it is not obviously safe either. The gap between the two judgments is exactly where money gets made or lost.
And then there is the weekend project that refuses to be boring. A developer has PS2 games statically recompiled to WebAssembly, which is a reminder of how much machinery sits just one layer above the metal. Rewriting PlayStation 2 titles for the browser is not a remake and not an emulator. It is compilation work that keeps the original code path intact while changing the runtime target. It also shows that modern compilers can still squeeze legacy targets to work inside a page, and that is a feat worth a look even if the games are already old.
Altogether, today is less about any single breakthrough than about the texture of a field moving in five directions at once. The models keep getting better, the tools for using them keep getting cheaper, the theory keeps getting sharper, the valuations keep getting louder, and the retro games keep getting runnable in a browser.
Sources
Every story from today's report, with its original link:
- twitter.com – Andrew Karpathy notes that public perception of large language models diverges from their actual capabilities, warning that misaligned expectations keep getting wider.
- tedhaley.ca – Ted Haley shares practical techniques for optimizing AI tool calling and cutting down token consumption.
- teorth.github.io – Terence Tao presents advanced mathematical research with computational insights.
- www.cnbc.com – A financial analyst criticizes AI IPO projections, arguing that investor enthusiasm may be misplaced.
- github.com – A developer demonstrates how classic PlayStation 2 titles can be rewritten in WebAssembly for modern browsers.