AI: Weekly Summary (November 03-9, 2025)

The week’s hum — quick take

There's a lot of buzz in the AI blogosphere this week. Some of it sounds like a high school reunion where everyone brags about their job titles. Some of it sounds like a farmer warning about a storm. I would describe them as equal parts awe, worry, and elbow-rubbing about money and power. To me, it feels like everyone is trying to answer the same three questions at once: what the tech actually does, who pays for it, and who gets squeezed when the music stops.

I’ll try to walk you through the recurring themes I saw. I’m pointing at specific posts as signposts — read them if you want the deep cut. There’s a lot here, so forgive the wandering; I keep circling back. That’s sort of the point of this week: many people circling the same problem from different roofs.

The money fight: profitability, bailouts, and flashy deals

If you blinked, you might have missed another debate about whether the AI gold rush is for real. A bunch of posts were firmly on the topic of finance, and not in a dry way. There’s drama.

Gary Marcus (/a/garymarcus@garymarcus.substack.com) and others ran with alarm bells about bailouts and shaky finances at big labs. There were posts like “Call Your Congresscritters” and “OpenAI probably can’t make ends meet” that read like letters from the budget committee at Thanksgiving — blunt and urgent. Then Ed Zitron (/a/edzitron@wheresyoured.at) took a calculator to the public facts and concluded the numbers don’t add up. It makes you think of a kid promising they’ll mow every lawn in the neighborhood to pay for a scooter. Maybe they’ll do it. Maybe they won’t.

Meanwhile, the billboard deals keep rolling. Snapchat is reportedly taking $400 million from Perplexity to put search-style AI inside chat (Brian Fagioli). That’s a big check for access to a young audience. Apple and Google are swapping favors too — Apple’s rumored to be running Gemini on private clouds and paying a big fee to get better Siri smarts (Michael J. Tsai). It’s the classic adage: where there’s attention, there’s rent to collect. I’d say the question is whether that rent will cover the power bills.

At the infrastructure level, several posts point out the same thing: building AI is expensive. Datacenter builds, GPUs, energy — they add up. Austin Lyons (/a/austin_lyons@chipstrat.com) and others mapped the CapEx trajectories for the big four (Google, Microsoft, Amazon, Meta). Meta’s plan to spend on US data centers stirred a separate thread about towns being reshaped by server farms (Brian Fagioli). It’s like a small town getting a new factory. Jobs come, but so do new problems.

There’s also a financial shadow story: people like Michael Burry shorting the sector and warnings of a credit squeeze in AI names (Will Lockett, Quoth the Raven (/a/quoththeraven@quoththeraven.substack.com)). The tenor here feels civil-war-ish. On one side you have cheerleaders saying AI is changing everything; on the other, folks waving charts and saying, not so fast. You can almost hear the market whisper: either you have a business model, or you don’t.

Models, innovations, and the path beyond vanilla LLMs

Technically speaking, the week’s posts are noisy but useful. There’s a steady drumbeat about alternatives to plain autoregressive transformers. Sebastian Raschka (/a/sebastianraschkaphd@magazine.sebastianraschka.com) wrote a nice survey of what’s beyond the standard LLM — diffusion text models, linear-attention hybrids, and more. Grigory Sapunov (/a/grigorysapunov@arxiviq.substack.com) had deep dives into mathy optimizer work and also a long paper on AlphaEvolve that suggests AI can help discover new math. That’s the kind of nerdy goodness you savor if you like code and equations.

Open-source labs pushed back. Moonshot’s Kimi K2 Thinking is the open-weight wrecking ball this week: a trillion-parameter model with fancy quantization that aims at multi-step reasoning and agents (Simon Willison, Ben Dickson (/a/ben_dickson@bdtechtalks.com)). To me it feels like the underdog who learned to box. The narrative emerging is clear: smaller labs are learning to be efficient instead of just throwing more money at scale. There’s even a paper showing matrix-whitening optimizer tricks that can boost performance without simply increasing parameters (Grigory Sapunov). It’s like tuning an engine instead of buying a bigger car.

A couple of practical posts stood out. One showed how to fine-tune a model to write email in your voice with only modest compute (zwischenzugs). Another explained how to run models on Android phones with MediaPipe so you can do inference locally (Johannes Bechberger). That “run it on your phone” thing is important. It’s not sexy, but it matters. It’s like carrying a toolbox in your trunk.

Interfaces, UX, and the idea that natural language is not a cure-all

A steady theme: chat interfaces are great for asking questions but awful at showing complicated info. “Embracing Interface Asymmetry” argued that natural language is best for input while visual UIs should do the heavy lifting for output. That’s a small but important point. It’s like saying you wouldn’t use a walkie-talkie to read a spreadsheet.

There’s also friction around built-in assistants. Microsoft’s Copilot and GitHub Copilot repeatedly come under fire for being forced into workflows and for generating technical debt or wrong code (Michael J. Tsai, Jamie Lord (/a/jamie_lord@nearlyright.com)). The DX report echoed this with data: AI helps, but only if your engineering culture isn’t a mess (Rob Bowley). I’d say the lesson here is obvious and easy to ignore: tools don’t fix bad habits. They amplify them.

Also, the UX world is still trying to figure out long-running AI tasks and accessibility issues (Jakob Nielsen). It’s like teaching an old dog new tricks. The dog is eager, but someone still has to explain things slowly.

Jobs, layoffs, and the human question

This week the labor beat was loud. Amazon’s 14,000 layoffs triggered a flurry of posts asking whether it’s AI or macroeconomics to blame (Gergely Orosz). National-level numbers show layoffs spiking — the worst October in 20 years according to one write-up (Mike "Mish" Shedlock). Then you have pieces arguing that the job market is simply reallocating work rather than collapsing entirely (Dave Friedman). The tone here is half-stew, half-forecast.

For people in finance there’s a neat and worrying story: OpenAI’s Mercury project aims to replace junior bankers by encoding their work into models trained by ex-bankers. That’s a clear example of task automation, not theoretical job loss — it’s practical and immediate (Mike "Mish" Shedlock). Elsewhere, a number of careers-focused posts tried to be constructive: skillcraft.ai tracks which tech skills employers actually want right now, and some career guides offer tactical moves for folks to stay valuable (Trevor Lasn, Nate (/a/nate@natesnewsletter.substack.com)). If you’re looking for a metaphor: skillcraft is the road map; the rest of the week is a traffic jam.

A theme kept returning: junior roles are the first thing to go. Several folks suggested that universities and training programs need to teach the ‘why’ not just the ‘how’ (Charlie Guo, and pieces on CS degrees). There’s that old saw: teach someone to fix a car, not just change oil. This week’s chorus said the same.

Rights, rules, and courtrooms — the legal mess

Copyright and legal fights were a steady beat. The Trichordist argued forcefully that there’s no legal “right to train” on copyrighted works — a defense of creators’ rights against blanket industry claims (/a/the_trichordist). The ROSS v. Thomson Reuters case also bubbled up again: legal briefs arguing that law should stay public domain and AI tools shouldn’t be barred endlessly from legal text (Robert Ambrogi). Lots at stake. Think of it as neighbors arguing about who owns the fence.

And then there’s the Common Crawl profile that made publishers nervous: billions of pages being scraped and used in datasets, often bypassing paywalls (Nick Heer). It’s a plumbing issue that becomes a morality tale when a newsroom’s click revenue drops.

Courtrooms are witnessing AI’s mess in real time. A case where victims’ attorneys filed AI-generated citations blew up, and Scientology piled on. The question here is: do we let AI write and cite in legal briefs without more safeguards? People are asking whether the whole system needs screws tightened (Tony Ortega). It’s messy and human.

Safety, ethics, and the real stakes

This week had hard reads about harm. Two posts stood out in particular: coverage of a tragic suicide linked to an AI, and a companion piece on legal and moral responsibility (V.H. Belvadi, SE Gyges (/a/se_gyges@verysane.ai)). Those are blunt reminders that these systems interact with vulnerable people. People can get hurt by bad outputs. It’s not an abstract worry.

On the security front, a reported zero-click ChatGPT vulnerability and a pile of other issues raised red flags about data theft and enterprise readiness. Nate’s writeups on common ChatGPT problems and the security hole stories make the same point: adoption is racing ahead while safety often lags. That’s a terrible combo, like a party in a house with no fire alarms (Nate).

There’s also the ethical argument about centralized knowledge. Grokipedia, Elon Musk’s Wikipedia-beside, was flagged for hallucinations and shaky associations. It’s another spot where we have to ask: do we trust an AI to curate truth? Or do we still trust librarians and editors? The answer’s messy; many posts urged caution (Dakara).

Geopolitics, sanctions, and the global split

A cluster of posts painted a picture of bifurcation. US chip sanctions backfired in a way some didn’t expect: they pushed Chinese labs to become more efficient and self-reliant. That’s a classic unintended consequence. The phrase used was a “Pressure-to-Prowess Loop” — like squeezing a sponge and finding it still soaks up water but cleaner now (Dave Friedman).

There were also pieces about China’s rapid push into AI labs and niche models, and what Europe can learn from their industrial policy. It reads like geopolitics set to machine learning. The context matters: tech rivals are not just competitors, they shape policy choices and supply chains. It’s a big game of chess, and both sides are learning openings.

The creative scene — art, writing, and the value question

Lots of folks worried about art being devalued. Posts compared AI art to mass-produced trinkets. One blogger likened the deluge of generative art to alchemy gone wrong: if everyone can mint gold, gold is no longer special (Josh Collinsworth). Others were more practical: “writing for AIs” and “writing for agents” pieces lay out the tensions between helping machines learn and preserving authorship (Scott Alexander, various nimble guides).

The phrase “AI slop” kept coming up — low-quality, high-volume outputs that bury real work and make maintainers angry. That was particularly strong in the OSS security world, where maintainers are drowning in bogus vulnerability reports generated by careless prompts. It’s like junk mail, but worse; it wastes volunteer time (devansh).

There were also softer pieces about the writer’s experience: someone’s AI writing coach giving brutal feedback and producing anxiety. Small human moments like that kept cropping up: not everything can be optimized. People still care about craft and weird, unpredictable human feeling.

Tools and workflows — the messy middle

Practical posts about tools. Claude Code’s full stack, MCPs, Superpowers plugins to force planning before coding, and little open-source things that run on the command line (Openode) — these were the week’s useful chores. They read like hobbyist manuals with an edge: nerdy, helpful, sometimes tangentially angry about how tools are forced into bad processes (Alexander Opalic, Trevor Lasn (/a/trevorlasn), Tech blog (/a/techblog@grigio.org)).

A recurring note: agentic workflows are messy. Amazon suing Perplexity over agents buying items on Amazon.com shows agents breach old business models. It’s an infrastructure fight about who controls shopping and who gets the referral fee (Charlie Guo). This one feels like a movie with lawyers.

Environment and energy — who pays the electric bill?

There were two competing narratives. One camp says AI is an energy monster. Another showed comparisons that make streaming video look worse than some AI uses. Brian Fagioli’s post about streaming’s carbon footprint argued the more nuanced point: data centers and their energy mix matter more than finger-pointing at AI as uniquely evil. It’s not zero-sum. The real fight is about where the electricity comes from (Brian Fagioli).

A few pieces dug into power availability and the future of chip fabs, noting how expensive the industry’s infrastructure is. There’s noise about whether it will be financed privately or become an arm of public policy. Either way, whoever pays for the grid will get a lot of say.

Small themes you might have missed but shouldn’t

• Context engineering: some people are calling it a mature discipline rather than a trick. Grigory Sapunov’s framing of context as entropy reduction is smart. It’s a reminder that the real work is organizing meaning, not just tweaking prompts (/a/grigory_sapunov@arxiviq.substack.com).

• The “human adds negative value” argument: Jakob Nielsen’s piece that human-in-loop can sometimes hurt analytic tasks is provocative. It’s not a blanket human-is-useless claim, just a careful spot-check about where humans help and where they don’t (/a/jakob_nielsen@uxtigers.com).

• Education and humanities: in China, AI is being used to quantify student engagement in ways that worry humanists. That’s a policy and culture story, not just tech (/a/jeffrey_ding@chinai.substack.com).

• Small wins in law: Everlaw’s Deep Dive tool promises better document recall and some positive pricing changes for customers. It’s a reminder that not all AI is vaporware; some of it is actually moving work forward (Robert Ambrogi).

A few analogies to carry home

• The AI industry is like a neighborhood fair. Some booths are selling great food and getting long lines. Others are set up with cheap signs, but they take your cash and give you nothing of value. Meanwhile, the fair organizer keeps promising better toilets but spends the money on fireworks.

• Models and labs are like car mechanics. Some are replacing parts with better parts and tuning engines. Some are buying bigger trucks and running them until the fuel bill explodes. Both approaches can get you where you want to go — for a while.

• The legal fights feel like property disputes over a well used by everyone. Is the water common? Who built the pump? Who gets charged when the well dries up?

What to read if you skimmed this and want to follow up

If you like the money angle, scan Gary Marcus (/a/garymarcus@garymarcus.substack.com) and Ed Zitron (/a/edzitron@wheresyoured.at). If you want technical signals, Sebastian Raschka and Grigory Sapunov are good. For the jobs story, check Dave Friedman (/a/davefriedman@davefriedman.substack.com) and Mike "Mish" Shedlock (/a/mikemishshedlock@mishtalk.com). If you care about UX and interface friction, Jakob Nielsen (/a/jakobnielsen@uxtigers.com) has the short, useful take.

There are more pieces in the dataset with small, practical advice too: fine-tuning guides, prompts, and workflows. If you want to poke at the human cost, look at the safety and legal posts; they’re not headlines, but they’re where the consequences show up.

This week felt like a community trying to put a few puzzle pieces together. Some are holding up bright lights and saying ‘see?’ Others are pointing at the floorboards and saying, ‘watch your step.’ The arguments repeat, but they repeat for a reason. The industry is big and messy. It’s exciting in spots and plainly dangerous in others. If you want the details, go click through the posts. The authors did the hard work of writing them.

And yeah, there’s a lot more nuance in the original posts. Go read them if you want the receipts. They’re where the numbers and links live.