Hacker NewsFriday, October 2

Critics question adding MCP, codemode, Jev support and Anthropic cache warming to a "minimal" agent. Armin from Earendil explains they follow how models are trained on their native harnesses; others say a small team simply uses judgement.

One user laments RAM-hungry TS/Python agents and wants static Rust binaries. Replies note Earendil chose TS for easy bootstrapping without ruling out a port, and that TS aids extensibility, LLM familiarity and self-modification.

A user quit after others reverted "not walkable" edits on sidewalk-less roads. The developer removed that controversial quest for v64.0. Others said OSM tags must stay objective and legal, not comfort-based, since access tags affect routing. One suggested sticking to stale POIs.

Contributors report that KMP is stable, with only about 2k of 80k lines platform-specific and few KMP bugs. Compose can be unwieldy and churns often, and JetBrains IDE bugs are annoying. Kotlin and Compose feel similar to Swift and SwiftUI, and KMP can share anything from one module to a whole app.

Beta testers got stuck in quests and could only exit via a left-edge swipe, a known bug. The "Uh..." option and undo button were pointed out. Praise for the low barrier to entry was common. Every Door and Go Map!! were raised as alternatives, and some noted the TestFlight link was hard to find.

Some say for-profit apps profit from keeping users single, while a government can track actual marriages. A dating app founder disputes this as user behavior, not conspiracy. Others link the effort to Singapore's very low fertility rate and debate taxes and incentives.

Commenters explain the algorithm favors the proposing side while receivers get their worst stable match, using examples. Others question applying it to thousands of users needing full rankings; replies note any comparator or derived ordering works. Its use in medical residency matching is mentioned.

Many argue Anthropic can't object to Chinese labs distilling its models after training on uncompensated human work. Defenders cite the cost gap and IP logic, but critics say the parallel is exact and that the complaints expose a weak moat.

Proposed reasons include CCP strategy to commoditize LLMs as a complement to manufacturing, involution and overcapacity among many labs, soft power, and DeepSeek setting a precedent. Others reject the idea that all decisions are centrally directed.

Users report strong local performance, such as Qwen on a 5090. Some argue cheaper inference undermines OpenAI and Anthropic's revenue-growth story. Others insist people will always want the smartest frontier model, or that the labs already have high inference margins.

Many argue decision models are easy for anyone with ML experience: constrain an LLM's output or tweak post-training. Some call Jev clever marketing of low-hanging fruit. Others say the real value is intelligence plus data and evaluation, not the architecture.

Users question claims that Clef beats Jev, noting many "Jev-level" models fail at non-trivial tasks like playing games. Public benchmarks are seen as easy to overfit. Cloudflare reportedly had parallel in-house work and long-standing classifiers for bot detection.

Clef is open weights, not open source, so it can't be reproduced. At about 6x Jev's input price, it's cheaper to self-host; output costs are negligible for decisions. Commenters note Jev can't take priors or be fine-tuned, though its outputs can feed downstream classifiers.

Users praise Bloomberg's dense UI, comparing it to avionics and NASA displays. Many say modern UX favors new users over daily experts, citing hidden menus and poor car UIs. Others note such UIs need heavy training, and that employers with high turnover prefer easy-to-learn systems.

Commenters note the modern Terminal runs on a private Chromium fork. A claim that 1985 hardware still works was corrected: museum units only display output from a modern PC. Others mention an OCaml-based derivatives DSL, a custom Matthew Carter font, and open alternatives like bergoom.

Users lament the roughly $25k yearly cost and point to the NYPL terminal, public demos and trials as ways to try it. A new physical-keyboard login requirement drew criticism, but others argue Bloomberg's bond-data monopoly makes users stay, and many like the keyboard anyway.

Commenters say border agents everywhere have broad powers, making borders a prime place to collect data. Users dispute claims about the UK's powers and Poland's easy entry, and note that treatment depends heavily on nationality and luck.

Users suggest backing up and wiping phones before crossing, carrying travel phones, using decoy profiles, or using GrapheneOS in BFU state while staying silent. Others warn wiping may be treated as destroying evidence, and that handing over a clean device may go more smoothly.

Many quote the 4th Amendment and debate the vagueness of "unreasonable," court-made border exceptions and the disputed 100-mile zone. Stats show searches rising since 2012 (~0.013% of travelers). Some see targeting of critics; others call it long-standing bipartisan policy.

Users debate whether ending ChromeOS in 2034 breaks the 10-year promise. Many say buyers from 2024–2026 were rugpulled, getting only 8 years. Others counter that hardware support continues via migration to Googlebook OS, so the promise is technically kept.

Some theorize the rebrand escapes negative associations from cheap school Chromebooks. Skeptics doubt anyone would pay $1k for a Googlebook. Others cite Google's habit of abandoning products and call the name awful, though a few praise solid Chromebook hardware.

Commenters say school Chromebooks rarely survive 5–7 years, so the cutoff matters little in practice. Others raise privacy and contract lock-in concerns. Googlebook OS is described as Android with a sandboxed Linux VM, raising fears of a shift from open ChromiumOS to closed code.

Some argue sustained high prices invite new entrants and point to past DRAM price-fixing. Others say fabs cost billions and take 5-10 years to build, so caution about overbuilding is rational. They note shortages would create the same incentives anyway.

CXMT grew from 1% to about 10% of DRAM and could gain more while the big three stay supply-constrained. Sanctions limit its equipment, and some doubt it can meet global demand. Users complain about prices like $1k for 64GB DDR5 ECC.

Commenters are skeptical of a memory seller predicting tight supply, likening it to a shovel seller urging buyers. Many see AI-driven demand as speculative, citing Acer's CEO forecasting lower prices by 2027. Others think inference demand and open-weight models will keep demand high.

Commenters argue early cars resembled carriages, giving farriers time and adjacent skills to pivot. AI moves faster, targets knowledge work rather than muscle, and builds a general thinking machine. The CGP Grey "horses" quote raises whether new jobs will always appear.

Many say retraining is largely a myth. Displaced workers face steep pay cuts, experience requirements, no money for school, and frozen entry-level hiring. Some urge adaptability and adjacent roles, while others note that even fallback jobs like driving or security face automation.

New jobs may benefit society overall but not the individuals and families thrown into poverty, as with the Luddites. Debate covers whether consumers remain necessary, inequality, UBI, and whether the rich will need people at all. Others cite Jevons paradox as grounds for optimism.