Hacker NewsWednesday, October 7

Some see Large 4 as mediocre and late, trailing US labs and Chinese open models like GLM-5.3 and DeepSeek. Others praise its efficiency, cyber/vision scores, and value for EU sovereignty and trusted inference, though some doubt it is truly independent.

A poster asks why hyperscalers need huge clusters if Mistral trained on ~3,800 GPUs. Replies cite inference demand, many experimental runs, distillation from frontier models, larger future models, and a closed-vs-open quality gap.

Some argue revenue still grew and the loss reflects investment, currency effects and lending to subsidiaries, not doom. Others suspect cash burned on Junie, tokens and AI FOMO while the core IDEs stagnate. Several commenters dispute what "cash flow from investing" actually means.

Many say they rarely open IDEs now, using Claude Code or Codex and reviewing diffs in GitHub or VS Code. Skeptics warn that unreviewed agent code is lowering quality, citing bubble sorts and 200-file PRs, and say hands-on review still matters for real products.

Users criticize late, confusing AI efforts (Junie, Air, curated models, a separate subscription), sunsetting Upsource and Space, slowness, bugs and price hikes. Many think JetBrains could still win by integrating bring-your-own agents with its strong refactoring and navigation tools.

Many criticized the early-access signup with no weights or tech report, echoing "show me the weights." Others noted weights are promised this month, called the criticism caustic, and some argued the email-harvest approach was a marketing miss versus releasing first.

Commenters noted Beam is larger yet trails DeepSeek V4.1 Flash and Kimi K3. Defenders say its key selling point is being an American "open frontier" lab for government and Western clients, and that new entrants deserve encouragement; critics call its training inefficient.

The claim that the world-map grid puzzle was "a few days old" and couldn't be in training data was disputed, as it appeared on LessWrong in August 2025. Users called it sloppy and debated whether web search tools could invalidate such tests.

Some say copying signatures shows AI is mechanistic pattern-matching without understanding or introspection. Others ask how human brains differ. Critics note hypocrisy: AI is called human-like to defend copyright use, but a dumb copier when it copies signatures.

Many argue a human forging a cartoonist's signature would be sued, so AI vendors should be too. Debate covers commercial use, statutory damages, German personality rights, and trademark vs copyright. Some suggest cartoonists register trademarks and act through unions.

Users report manually erasing fake signatures. Others suggest automated review passes or negative prompts, though these add cost. Some blame users who share images without checking, while others say AI tools are designed to encourage lazy, one-and-done use.

Several users flagged the post as a dupe of an earlier submission that already had about 200 comments. A moderator confirmed the comments were moved to that thread.

One user asked whether the item had been released again. A reply, apparently sarcastic, noted it was "clearly a different URL," suggesting the same content was resubmitted under a new link.

Users explain neutrinos as "ghost particles" detected via Cherenkov radiation in deep Antarctic ice, clarifying that particles only exceed light's speed *in the medium*. They also cover why Antarctica (stable, transparent deep ice, existing logistics) and how directions reveal cosmic sources.

Many romanticize remote work, while former IceCube staff share experiences: Linux admin duties, wintering, boredom after two weeks, and the "adventure, money, misfit" joke. Others note the heavy logistics of the South Pole Traverse and air cargo.

Commenters note the rare solo physics award and debate whether others deserved credit, given IceCube's 400+ collaborators. Some argue Halzen fits as UW-Madison PI who conceived and actually made the hard-to-fund project happen; a former student praises him as decent.

A magnetic materials PhD calls the ferromagnet/antiferromagnet binary bizarre, ignoring diamagnets and paramagnets. Others note domains, Halbach arrays and ferrimagnetism, suspecting unreviewed AI writing and an author unable to validate results.

Many readers misread "semiconductor" as "superconductor," evoking LK-99. Commenters note room-temperature semiconductors are ordinary and that "antiferromagnetic" was dropped from the title. They debate whether LK-99 was a hype debacle or science working.

The agents ran standard DFT simulations (Quantum Espresso), essentially LLM-driven search against traditional evaluation. One material existed in a 1999 paper, and vals.ai is seen as hype-driven. Synthesis and testing, not candidate generation, is the real bottleneck.

Many call Polars a near-full Pandas replacement with a cleaner API and huge speedups. Holdouts cite familiarity, interoperability, row-wise operations, and missing GeoPandas parity (GeoPolars is immature). Some debated nanosecond timestamps and string concatenation.

Users compared Polars with Postgres/SQLite, noting its in-memory columnar design suits analytics over OLTP. DuckDB is favored for huge joins, lower memory use, and geospatial work. New SQL support is seen as competing with DuckDB, and some asked why Polars and DataFusion don't converge.

A benchmarking veteran warned against reading release benchmarks as absolute rankings, since TPC suites lack diversity (e.g. JSON workloads). A Polars developer described fair out-of-the-box methodology, overhead on 192-core machines, and planned join fixes. Others reported memory/OOM issues tied to threading.

Some argue the complaints are overblown: the VM is user-owned, contact profiles stay private, and Messages access required Full Disk Access. Others cite Meta's privacy record, encryption doubts, and third-party data exposure.

Many refuse to give agents bank, email, or chat access due to LLM unreliability and counterparty risk. Suggested mitigations include limited-fund cards, purchase confirmations, vendor liability, and self-hosting. Others say most users won't care.

Commenters expect ad giants to eventually exploit intimate assistant data to sell products or manipulate behavior, predicting enshittification. Others note Meta already profiles everyone, so this is business as usual.

Many wondered how many tests broke from the redesign, citing Hyrum's Law and IANA's warning not to rely on it. Some suggested randomizing the page; others noted AI agents now write such tests, and a self-hostable replica of the classic page was offered.

Users rely on example.com's plain HTTP to trigger Wi-Fi login portals. Alternatives like neverssl.com and captive.apple.com were suggested, while pinging anycast 1.1.1.1 was criticized as an unreliable connectivity test.

The rotating translation animation was dropped for accessibility, now listing all languages with the user's first. Others joked about its loss as a white glasses-cleaning screen, its minified HTML, per-letter spans, and the article's unlinked mentions.