Today’s Highlights#

The day’s biggest story came from Europe. Mistral Large 4 opened its public preview: a 1 trillion-parameter multimodal model whose weights are due to ship by the end of the month. Mistral’s pitch is sovereignty as much as benchmarks — it trained the model from scratch in its own European data centers, and buyers can run it on their own hardware instead of depending on a supplier on the other side of the Atlantic.

That framing is also the through-line of today’s edition. The AI race is no longer only about who has the smartest model; it is about who can actually run and govern one. OpenAI published a batch of mathematics produced by an internal frontier model and started negotiating with researchers over how machine proofs should be released. At the other end of the spectrum, Meta’s AI agent Muse became a case study in what goes wrong when an assistant ships before it is ready. Beyond AI, the Nobel Prize in Physics went to the scientist behind the IceCube neutrino observatory, a Czech developer-tool company grew revenue while losing money, and a geothermal plant in Utah beat its schedule. Rounding things out, the internet’s oldest placeholder page got its biggest redesign in decades. Here is the tour, by topic.

Tech and Products#

Mistral Large 4 puts European AI sovereignty to the test#

Mistral’s preview release describes a 1 trillion-parameter, natively multimodal model with 49 billion active parameters — the “active” fraction that fires on each request, which keeps serving costs down. The company claims state-of-the-art results among open models in cybersecurity, finance and law, with some visual tasks ahead of closed frontier models. Training ran on 3,800 NVIDIA Grace Blackwell chips in Mistral’s own European data centers, across data in more than 160 languages. Weights go open at month’s end; in the meantime customers can self-host, which Mistral frames as protection against a provider cutting access mid-incident. The post also puts numbers on its reinforcement-learning pipeline and calls this the first milestone of a €3 billion Series D.

Note: the interesting part for a general reader is not the parameter count but that “capability you can hold yourself” has become a geopolitical option. The source, naturally, only reports results that flatter it, so some discount applies. The thread split over whether catch-up is still possible: several commenters argued the early years of AI have not turned into a winner-take-all race and that Mistral is still on the board after a rough previous release, crediting open Chinese labs’ papers and ordinary job-hopping for keeping the field porous. Others countered that compute and ecosystem, not model quality, are the real moat, and a few readers warned that the published comparisons lean on open models from months ago. Surprisingly many focused on the modest 3,800-chip cluster, calling it remarkable to get near the frontier on hardware that small.

Discussion: Hacker News thread

Polars 2.0 changes the default answer for tabular data#

The Rust data-frame library Polars shipped 2.0, and the headline feature is spill-to-disk: workloads that no longer fit in memory now overflow to the disk instead of crashing. SQL, the longstanding lingua franca for asking questions of data, becomes a first-class interface, and the project’s own benchmarks on the standard TPC-H and TPC-DS suites put it ahead of DuckDB and DataFusion. The release also adds a Map data type and stricter type checking so mistakes surface earlier, which matters as more code is written by AI tools. To its credit, the team admits a fixed overhead on very large machines running small queries and publishes a full benchmark repository for anyone who wants to check the numbers.

Note: the Polars-versus-pandas relationship is roughly automatic versus manual transmission — for everyday data work the difference rarely changes the result, but switching still costs effort. That was the comment section’s central question: whether Polars is now a full pandas replacement. Most readers said yes except for geospatial work, where a Polars equivalent of GeoPandas is still immature, and several noted the API is simply easier to reason about. A frequently shared comparison had pandas taking over four minutes on a billion-row challenge against about five seconds for Polars and DuckDB. Not everyone was convinced: some prefer DuckDB or the pandas ecosystem and called third-party library compatibility the real migration barrier, while one developer flagged edge cases like exporting nanosecond timestamps.

Discussion: Hacker News thread

example.com gets its biggest redesign in decades#

example.com is the reserved domain the Internet Assigned Numbers Authority (IANA) has kept for documentation since the late 1990s, the address you paste into a manual when you need a URL that means nothing. DebugBear’s write-up walks through the September 28 overhaul: the static English page became a JavaScript animation cycling through six languages every five seconds, each character fading in with its own delay, before the animation was dropped again on October 3 in favor of showing all languages at once. IANA’s statement says the content was split into a separate script file to cut bandwidth, since most visitors are automated and never fetch it, and reiterates that the domain is not a service and should not be used for availability testing. The article also replays the page’s look back to 1999.

Note: a placeholder nobody is supposed to visit now needs bandwidth optimization, which is a fair snapshot of how the web actually runs. One commenter discovered the change had broken their tests and offered a pixel-perfect replica of the old design, but the larger share of the thread asked why any test would depend on the page’s contents, sliding into the well-known rule about behaviors others depend on. A network engineer made the opposite case: example.com speaks plain HTTP, is tiny, memorable and essentially permanent, which makes it close to a perfect “can I reach the internet” probe, especially for forcing open a hotel-style login portal. Others cited the standards documents to argue that relying on its HTTP service was never sanctioned, and one reader joked the page’s business model is about to be disrupted by an AI startup.

Discussion: Hacker News thread

Business and Platforms#

JetBrains posts record revenue and still loses money#

Data from Czech company registry aggregator Helgi Library shows JetBrains turned over CZK 16.0 billion in 2025, up 6.3% and a company record, making it the largest of 15 media companies in the country by size. The same year ended with a net loss of CZK 315 million and a return on equity of -11.1%, a sharp swing from the prior year. The income statement behind the post shows gross profit sliding from roughly CZK 5.3 billion in 2021 to CZK 1.7 billion in 2025 while operating expenses climbed, pushing operating profit negative. The submission itself is little more than a table of figures; the substance sits in the comments.

Note: a software company with no fundraising story and a pure subscription model losing money is a clean specimen for anyone wondering whether the software business still works. The thread offered competing explanations: some readers assumed AI compute, salaries and hardware inflation ate the margin, while others gave a short accounting lesson on why investment spending is not the same as losing money and complained that markets only watch the profit line. One commenter went to the free Czech corporate registry filing and argued the loss is mostly currency movement between koruna and dollars in 2025, with group-level revenue actually up about 25% — “definitely not an AI-doom scenario.” A developer added a more ominous data point from the demand side: having not opened a JetBrains IDE in six months, they expect to cancel next year, a small sign that AI coding tools are reshaping what developers will pay for.

Discussion: Hacker News thread

First enhanced geothermal plant completes in 23 months#

Fervo Energy announced that its Cape Station plant in Utah started selling power to the grid on September 30, a day ahead of schedule and 23 months after groundbreaking. Enhanced geothermal means applying oil-and-gas drilling techniques to deeper, hotter rock, so a well no longer needs a natural hot-spring neighborhood above it. The first block delivers about 33 megawatts of a planned 100, with the site holding potential for as much as 4 gigawatts; Google and Southern California Edison have already contracted to buy power. Fervo went public in May, raising $1.9 billion on top of more than $1.3 billion in earlier venture funding, and wants future blocks in 18 months to serve data centers that need electricity fast.

Note: for a general reader the stakes are simple — always-on clean power just moved closer to reality, and data-center demand is deciding where the next plants get built. The comment section was not impressed. One reader pointed to a list of more than 50 enhanced geothermal projects dating back to the 1970s and questioned the “world’s first” claim, noting the plant actually came online at 33 megawatts rather than the 100 in the headline. The sharper cut was cost: roughly $7,000 per kilowatt for the first phase against about $2,157 for a gas plant, which led several readers to argue solar plus batteries plus some wind already wins by a wide margin — one quipping that the panels got cheaper again while you read the article.

Discussion: Hacker News thread

Policy and Governance#

A $111B merger closes and leaves two media giants in one#

Ars Technica reports that Paramount Skydance completed its $111 billion acquisition of Warner Bros. Discovery, forming a company called Skydance that now houses two major movie studios, the Paramount+ and HBO Max streaming services, CBS, CNN and a large sports-rights library. The deal had been challenged by California and 11 other states, and a federal judge had ruled the combination was likely to reduce competition substantially; California settled last month and the other states followed, while the Supreme Court denied an emergency filing from five consumers without comment. Among the settlement terms is an editorial independence board for CBS News and CNN, appointed by Skydance and reporting to its board. The Justice Department approved the deal in June, and the FCC allowed the company to finance it by selling non-voting stakes to the sovereign wealth funds of Saudi Arabia, the UAE and Qatar. The merged company is expected to carry about $80 billion in debt, with the Ellison family and RedBird holding all voting shares.

Note: who owns the newsrooms is a public question, which is why this belongs in a digest for general readers — editorial independence was written into a court settlement for the first time. The thread split sharply. One camp worried about accelerating media concentration and argued over the owners’ politics; another pushed back that such language risks sliding into stereotypes about Jewish people and asked that criticism stay anchored to verifiable conduct. Some pinned their hopes on a future administration reopening antitrust review, while more commenters doubted regulators would act at all. Setting politics aside, several readers did the arithmetic instead: $80 billion of debt means the company has to defend streaming growth, cable cash flow and theatrical releases at the same time, and any one leg failing will hurt.

Discussion: Hacker News thread

Science and Research#

The 2026 Nobel Prize in Physics goes to Francis Halzen#

The Nobel Committee announced that the 2026 physics prize is awarded solely to Francis Halzen of the University of Wisconsin “for decisive contributions to the IceCube Neutrino Observatory and the discovery of high-energy neutrinos of astrophysical origin.” IceCube sits at the South Pole and uses a cubic kilometer of clear ice as its detector: boreholes drilled 2.5 kilometers down hold roughly 5,000 optical sensors that wait for the faint blue flash left when a neutrino finally collides with an atomic nucleus. That flash, called Cherenkov radiation, is the glow a particle makes when it travels faster than light does in that medium. Neutrinos barely touch matter, so they cross the universe untouched and carry information straight from cosmic cataclysms, which is why astronomers treat them as a new kind of messenger.

Note: turning an invisible particle into front-page science took an engineering miracle that turned a block of Antarctica into an instrument. The comments read like a crash course: one reader gave a plain-language walkthrough of how the detector works, and another quickly corrected a common misreading — the “faster than light” bit means faster than light travels in ice, not in vacuum. A newcomer asked why the origin of the neutrinos matters, and a commenter with astronomy background replied that IceCube was never built to study neutrinos as such; it studies the universe through them, and it can now pin down the Milky Way and several distant galaxies as sources. A side debate covered why ice rather than water, with one reader noting the same blue glow is what you see in the cooling pools of nuclear reactors.

Discussion: Hacker News thread

OpenAI shares mathematics produced by an internal model#

OpenAI says it is releasing a batch of mathematical results produced by an internal frontier model, following advice from the Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study on how such work should be shared. The results land in a GitHub repository with protocols for revisions and citations, and many proofs come with Lean formalizations — machine-checkable versions a computer can verify step by step. For transparency the company also published ten summaries of the model’s reasoning, the compute spent (about three hours of ChatGPT Pro thinking per result on average) and counts of attempted problems. OpenAI says it will fund workshops around major results and is working toward responsibly releasing the model itself.

Note: the genuinely newsworthy part is not any single theorem but that researchers are now writing the rulebook for citing machine proofs, which will decide how soon the rest of us can check this work. The thread quickly became a debate about whether AI is hitting an inflection point. One self-described worrier asked the optimists what they expect over the next one to five years; a common reply was “use AI for your actual work and you will see the limits,” which drew the counter that a few years ago models could not count the letters in a word, and they are only getting cheaper and stronger. One reader claimed to have read several preprints and called the approaches genuinely novel, and was immediately challenged on how that judgment was possible within an hour. A more measured voice argued mathematics is an unusually verifiable field — formal proofs plus third-party review lower the risk — while stopping well short of assuming unlimited exponential growth.

Discussion: Hacker News thread

Society and Culture#

Meta’s Muse keeps tripping over privacy and security#

TechDirt rounds up the rough weeks since Meta launched Muse, its consumer AI agent: the product shipped with a zero-day flaw that could be used to spy on Mac users; a YouTuber who handed it their Facebook Marketplace account watched it sell belongings far too cheaply and hand out their home address; researchers found that pretending to be a Muse agent was enough to gain root access on the device. The agent also reads private messages without approval and uploads them to the cloud, which forced Apple to change macOS permission settings for full disk access. Reporting cited from Wired says Muse builds detailed profiles of friends and family, and 404 Media reported internal pressure to patch vulnerabilities late so the launch would not slip, with engineers quoted expecting an eventual large breach. The author’s read: regulators are too weakened to slow a land grab.

Note: letting one agent book tables, pay bills and read your messages is handing a beta product the keys to your house, which is why this story reaches past the tech press. The comments mostly reached the same conclusion: without financial liability nobody will hand over a bank account, an inbox or a chat history. One thread proposed that vendors would have to guarantee full reimbursement before trust becomes possible; another argued purchases need a hard switch that requires a human to click yes. A dissenting voice reminded the thread that Hacker News is not a representative sample and that Meta’s billions of users may simply not care, while a few commenters shifted the frame entirely, saying the missing piece is an insurance and accountability layer — and that companies will just price the risk into the service.

Discussion: Hacker News thread

Alan Kay’s early history of Smalltalk, published in 1993#

A long paper by Turing Award winner Alan Kay, “The Early History of Smalltalk”, has been put online as readable HTML with the original scanned PDF and a repository for corrections. Written for a 1993 conference on the history of programming languages, it traces the 1960s ARPA community’s ideas about human-computer symbiosis through Sketchpad, Simula, Doug Engelbart’s NLS and the Xerox PARC work that produced overlapping windows, iconic programming, byte codes and programming for children, ending with the Smalltalk-80 release. Kay is insistent that his colleagues were not trying to build a better mainframe: they wanted computing to work like a Montessori classroom, a learning environment for millions, and he casts the shift as a paradigm change on the scale of the printing press.

Note: the windows and icons on everyday devices inherited their grammar from this moment, which makes the paper worth reading as cultural history rather than a technical artifact. The thread mapped the descendants. One commenter showed how deeply Smalltalk shaped NeXTSTEP and Objective-C, visible in Apple’s development tools to this day, and another extended the family tree through Mesa, Cedar, Modula-2 and Plan 9 all the way to the just-in-time compilers inside today’s browsers. The historically minded pushed back on the hagiography: Smalltalk-76 only adopted the Simula 67 object model, a fundamental break from earlier versions, so Kay’s recollections should not be the only source consulted. Several readers ended on the same wistful note — a lineage this influential, and almost nobody has actually used the language itself.

Discussion: Hacker News thread

Closing#

Today’s edition opens with a model trained in Europe and closes with a conference paper from 1993, with South Pole ice, Utah drill rigs, a federal settlement and a Czech balance sheet in between. The technology keeps moving, and so do the arguments around it: who owns the capability, who absorbs the risk, how results get published, and where the old knowledge lives. None of that is exclusively an engineer’s business. See you next time.