Mistral releases the weights of Mistral Large 4: 1 trillion parameters, A52B, and an API price cut of nearly half (1308 points / 360 comments)
Mistral announcement · Weights on Hugging Face · HN discussion
On Oct 29 Mistral put the weights of its flagship Mistral Large 4 (codename Le Chonk) on Hugging Face (Mistral-Large-4-1T-A52B plus an NVFP4 build) — a 1-trillion-parameter, natively multimodal model that entered public preview a month ago; after three more weeks of RL it leads open-weight coding and defensive cybersecurity, and the API price drops from the preview's $1.36 in / $4.18 out per 1M tokens to $0.7 / $2.1.
WHY IT MATTERSThis directly resets what you pay for a frontier-class model. The hard numbers: 19.7% on Cognition's private FrontierCode 2.0 (which measures whether a repository maintainer would merge the agent's change), leading open-weight models on FrontierCode and Terminal-Bench, matching the best open-weight score on the Artificial Analysis Coding Agent Index with better token efficiency, and a 96.6% refusal rate on malicious cyber requests — which commenter [3form] immediately counters: "where is the statistic about not refusing the non-malicious requests?"
"Open" here needs an asterisk. As [rcr-anti] points out, it ships under a modified MIT license that is not OSI-approved — "if you're serving it as a business you need a license." Public weights are not the same as free commercial use: self-hosting is effectively open, selling it as an API is not.
The real barrier is hardware. At 1T parameters with A52B active, the model is far beyond a single machine: [jamesblonde] has a 4×H200 box with 540GB of memory and finds only 4-bit or even 3-bit will serve it. So "open weights" in practice means accepting quantization loss, or renting a room full of GPUs. And [d2kx] shows the realistic usage pattern: an EU-based B2B company runs Mistral Large 4 alongside GLM 5.3 Flash and DeepSeek v4.1 Flash, and says that a few months ago they weren't sure which EU provider to pick — now they're happy. What to do: if data residency and compliance matter, this is the "EU provenance + self-hosted" option; if only cost and quality matter, wait for real-world quantized numbers, not the leaderboard.
Google loses $500B in a day: AI summaries are eating its own search ads (50 points / 23 comments)
BBC · Ars Technica: AI search makes less, so more ads · HN discussion
On Oct 29 the BBC reported Alphabet's market value fell from $3.97T to $3.46T (about -$500B) and the stock dropped nearly 13% to a six-month low; the quarter itself looked fine (revenue $127.6B, +25% YoY; Cloud $31B, +103%), but profit fell 25% and free cash flow hit -$9.06B — only the second negative quarter in its history and nearly double the previous one — with management pointing at AI summaries squeezing ad slots.
WHY IT MATTERSThis is the first hard number for "AI eating its own cash cow," not a mood. Search ads have long been most of Alphabet's revenue; search now returns AI summaries by default, so there are fewer places to sell ads — and [refurb] cuts through the headline: "It's hilarious to see 'AI fears' in the headline when the truth is Google missed its earning estimate." Same day, a historic figure went largely unnoticed: Google Search fell below 50% of Alphabet's total revenue for the first time.
Two implications for you. First, the subsidy era of "free AI search" is ending: Ars Technica reported the same day that Google admits AI search makes less money and is fixing it with more ads — expect denser ad slots, not better answers. Second, AI infrastructure is now being priced by earnings rather than by narrative: last issue we logged a shelved data-center bond sale and China's STAR50 down 40%; this is the third data point — even the richest player's free cash flow has turned negative. What to do: if your product depends on frontier APIs, put price rises and quota cuts into next year's cost model instead of treating today's prices as permanent.
Two veterans of software warned the same day: don't let AI write your prose, and don't let it carry production responsibility (407 points / 212 comments + 89 points / 131 comments)
Martin Fowler: LLM Voice · Rachel Laycock: Why good developers are wary · HN: LLM Voice · HN: agentic
Martin Fowler published LLM Voice on Oct 29 — the feeling, while reading, that a model wrote this, and the urge to stop reading that follows; the same day, Thoughtworks CTO Rachel Laycock asked why good developers are so wary of agentic development, and answered: "built" is not "working in production."
WHY IT MATTERSTogether these are the most practical "usage boundaries" you'll get this week. Fowler's rules are concrete: do not let an LLM write the first draft — even when a human edits afterwards, the LLM voice is very hard to expel (the one exception he's seen needed a professional editor at a major publication to rewrite it); using a model for an outline or structure is acceptable, but authors should write the prose themselves; even accepting rephrasing suggestions is risky. His own use is narrow: AI as a search tool, with every fact verified against the source.
Laycock turns the wariness into an engineering question. Good developers aren't afraid of losing their jobs or of giving up typing — they know early coding agents made it easy to produce code nobody would want anywhere near production, written by people who haven't learned the chasm between building something and running it safely. Their actual questions: Is this the right behaviour? Does it work as expected? Are these the right boundaries? Will we be able to evolve it? What happens when it fails? Can another team live with it? Underneath all of it: how much can we trust a system when nobody has experience of building it?
Two actionable conclusions: ① for text that humans read — docs, external comms, proposals — don't let a model start the draft; the time you save comes back as a trust discount (commenter [PaulKeeble] is blunter: LLM text has "really low information density" and he can't stay interested); ② for agent-written code, define acceptance as "who maintains it and how we roll back," not "it runs." Neither rule is new, but today two people with decades of experience wrote them down as explicit policy.
Jensen Huang's "close relationship" with an alleged Chinese spy — and what the report does not say (310 points / 216 comments)
New York Post (relaying NYT) · HN discussion
The NYT reported on Oct 29 (relayed by the New York Post) that US intelligence concluded Chinese citizen Alice Huang (Huang Le) was passing information to Beijing, and that she had a "close relationship" with Nvidia CEO Jensen Huang; surveillance identified him as one of her contacts, but the report offers no evidence he knew about her alleged intelligence work, nor what she passed or how valuable it was.
WHY IT MATTERSThe tabloid surface hides the real story: how chip export controls get circumvented. Alice Huang is a top executive at Singapore-based Megaspeed, an Nvidia customer and cloud partner that provides AI compute to Chinese companies including Alibaba; US authorities are investigating whether Megaspeed helped Chinese firms evade export restrictions and move advanced Nvidia chips into China — buying through the US arm of Inspur and renting compute from Southeast Asian data centers to Chinese customers.
Three practical implications: ① compute compliance is moving from "do you sell the chips" to "do you rent the compute" — if your team buys overseas compute through third parties, this is the link now under scrutiny; ② "who knows whom" is becoming part of supply-chain risk, especially at the executive level; ③ keep the skepticism too: [pr337h4m] calls it "an apparent nothingburger" and notes "close relationship" is "a really good weasel word," while [tehjoker] asks whether CEOs talking to politicians counts as well — on the facts, the claim currently rests on intelligence findings, not public evidence.
"Zombie stats": when you chase the numbers AI cites, the source often isn't there (155 points / 66 comments)
Trakkr research · HN discussion
Trakkr ran a sourcing experiment — 200 everyday questions that call for a number, put to ChatGPT, Perplexity, Gemini, Claude and Google AI Overviews — then clicked the citations and kept clicking; many statistics, at the end of the chain, have no underlying study and turn out to be self-referential.
WHY IT MATTERSThis changes how you (and your boss) should use AI for data. The example: asked what percentage of salespeople give up after one follow-up, Google AI Overviews answers "44%" — and that number traces back through decks, blog posts and reports to no original research at all, a "zombie statistic" that looks authoritative because it is cited everywhere. The thread's number: about 40% of AI-quoted statistics do trace to a real study ([falcor84] deadpans that "73.6% of all statistics are made up on the spot"), which leaves roughly six in ten that don't survive the trail.
What to do: ① any hard number — percentage, dollar figure, ranking — must be traced to a primary source, otherwise treat it as unverified; ② prefer tools that support sentence-level citations (commenter [alberth] singles out Perplexity's "every sentence must be referenceable" design as a trust superpower); ③ keep [zzzeek]'s point: people who say AI should never be taught in schools are wrong — being correctly critical of LLM output is a huge skill, and it applies to human writing too. And note [jrflo]'s caveat that the report is a thinly veiled ad for the vendor behind it — hold some skepticism about the conclusion itself.
A privacy company turned "embrace AI" into a loyalty test: inside Proton's reckoning (148 points / 53 comments)
404 Media investigation · HN discussion
A 404 Media investigation published Oct 29 reports that Proton, the Swiss encrypted-email company, is pushing AI with unusual force: when an employee in the security Slack channel suggested spending a few days pentesting newly written code before shipping it, CEO Andy Yen replied that this was "bullshit and loser thinking" — "only losers think like that" — and told the team to "stop lovingly handcrafting all our code and use AI more."
WHY IT MATTERSThis is a complete sample of how an "AI transformation" destroys organizational trust. The reported measures are specific: an email address for employees to flag anything blocking AI adoption so leadership can "ruthlessly smash everything that stands in the way of AI velocity"; software installed on workers' computers to track how much AI-generated code the company produces; the CEO saying this is the last year Proton will hire engineers at all, only people who build things with AI; and this week, re-interviews of staff about their AI use, with the clear implication that it could decide whether they stay. 404 Media spoke to eight current and one former employee and reviewed Slack logs, meeting recordings and a CTO-written AI manifesto, describing a climate where security concerns can't be voiced; an internal document attributes missing the annual recurring revenue target to "insufficient velocity."
The point isn't schadenfreude, it's pattern recognition: ① making AI usage a KPI suppresses quality objections — exactly the objections a security product most needs; ② users are already voting with their wallets: [didacusc] says "just renewed my two year sub. Auto-renewal to be switched off imminently," [Refreeze5224] says "I no longer trust Proton after reading this" and is considering migrating to Tuta, and [pier25], a long-time customer, says "other than some UX improvements I'm not missing any big features. This idea they need more velocity is a failure in management"; ③ if your own company is doing something similar (tracking AI-code share, treating adoption as attitude), treat this as an early warning: the first people to leave are usually the ones who spot the problems.