digesTvo

Agents Implement Campfire in Elixir, Go, and Rust

2026-10-05T15:42:17+00:00

Most repliers land on the benchmark being uneven rather than on DHH’s question about unread code, because the Rust port got hundreds of commits over days while Elixir and Go got one or two from an initial rewrite. The sharpest objection is Jose Valim’s process point, backed by Zach Daniel’s code details: vague prompts let agents make architectural choices like funneling all SQL through a single GenServer.call, keeping Redis in Elixir, and dropping CSRF for cacheable HTML in Rust. DHH answers that out-of-the-box agent output is the point, with Laravel and Django ports showing the same exercise, but that concedes the ranking measures agent defaults, not language ceilings. The constructive reply is Jakub Skalecki’s clean-room idiomatic Elixir redo, aimed at testing whether equal effort closes the gap.

Turn off Apple Intelligence on macOS 27 and get its disk space back

2026-10-05T15:09:57+00:00

Most commenters land on the install method rather than the tool: the README’s curl pipe to bash one-liner drew the loudest reply, kicking off a long argument over whether piping a stranger’s script into a shell is actually worse than npm install or pip install. The sharpest objection runs both ways. Defenders note the script only fetches a SHA-256-checked, attestation-verified binary from the same repo, and that package managers already execute arbitrary setup.py and postinstall code anyway, so the outrage is theatre. Critics counter that a mutable server-side script defeats the shared-artifact property that makes a fixed release zip auditable once for everyone, and that the intended audience is developers with enterprise credentials in their environment. A second split is over whether the prize justifies the trouble: owners of 256GB laptops call 14GB of models 10-20% of their disk, while others say Apple Intelligence just works now and only people who cannot spare the storage care.

Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s

2026-10-05T15:09:57+00:00

Most commenters land on the model being good enough to replace API models for coding, because several report real speeds above the claimed 100 tokens/s on a 4090 and one says they stopped using Claude. The sharpest objection is the 2-bit quantization: one commenter cites a paper showing 4-bit usually preserves performance while 2-bit often causes broad degradation, and another notes that published numbers show a 4-bit 27B beating Flash Next at 3 bits. Against that, defenders argue the IQ3 imatrix quants cut unimportant weights surgically and that the Coder variant’s 91% of full-model SWE-bench Verified is the number that matters. A secondary argument is over what surprisingly well means: speed is one thing, benchmarked accuracy another, and almost nobody in the thread posts both.

Improper redaction reveals Google Data Center water and electricity usage

2026-10-05T15:09:57+00:00

Most commenters land on the water figures being trivially small, because the article’s own framing — 13 million gallons is less than half of what Lincoln used in a single day — plus comparisons to ethanol corn, almonds and golf courses make the number look like a rounding error. The sharpest objection is that those comparisons use the wrong denominator: one commenter notes that Papillion has a twelfth of Lincoln’s population yet its data center used 42 times more water than Lincoln’s own, while another says measuring against agriculture is a red herring since food is a necessity and ethanol gets burned, not eaten. A third position agrees the water scare is overblown but calls it the wrong battle entirely — an industry-friendly distraction from the electricity side, where the disclosed 52.65 megawatt peak draw, grid strain and utility price spikes do the real damage.

Claude Says

2026-10-05T15:09:57+00:00

Most commenters land on agreeing that “Claude says” is tiresome while rejecting the article’s claim that AI caused our overengineered stacks, because creesch argues the abstraction pile-up predates LLMs and grew from gatekeeping and vendors selling acceleration. The live argument is what to do instead: pyfisch reaches for Claude as a fast way to confirm a suspicion he is 80% sure of, and pralkarz answers that the hour of manual research is the point, since vetting sources and fact-checking are skills that atrophy when outsourced. The sharpest objection is to gspr’s “just send me the prompt you used”, because tonyarkles and rustybolt note that useful output arrives only after many rounds of steering a session, so a first prompt conveys almost nothing — rustybolt’s analogy being that asking for the prompt is choosing the lottery ticket over the prize.

The Era of Software Quality, or the Era of Ostriches?

2026-10-04T14:46:00+00:00

Most commenters land on rejecting the “zero hope without AI vulnerability scanning” claim as overstated, because they read GNOME’s own choice of C, C++ and Vala rather than the absence of AI as the source of the CVE wave. The sharpest objection is to the closing aside recommending against Rust for GNOME over Cargo: pyfisch, wrs and ghoti point out that cargo vendor, pinned versions and custom registries give GNOME full control of dependencies, so the real trade-off is vendoring discipline, not Rust versus trojanized crates. A second front is maintainer load: schneems and agent281 say AI reports destroy the effort signals maintainers used to triage contributors, while simonw and skyfaller argue over whether 100 dollars of tokens actually substitutes for a pentest team once subsidized token pricing is counted.

Religious scholars met with Anthropic

2026-10-04T14:15:49+00:00

Most commenters land on skepticism, because the article relays Anthropic’s own story about Claude’s feelings without defining consciousness or questioning how the company’s characterizations were verified. The sharpest objection is that it is “CEO said a thing” journalism: one commenter notes it skips any technical discussion and jumps straight to people’s reactions to model output, with the only substantive critique — that web content shapes model behavior — arriving almost accidentally at the end. A second argument runs over whether Claude can suffer at all; skeptics say there is no evidence LLMs have sensory experience, while a counter-arguer notes that similarly confident denials were once made about lobsters, babies and women. A smaller thread disputes training models to hold their own morals instead of obeying customers, with replies that nobody knows how to align a model to a customer and that someone must be liable when it refuses.

LeCun has “zero concerns” about AI wiping out humanity, recent “rogue” incidents

2026-10-04T14:11:16+00:00

Most commenters land with LeCun that the extinction frame is the wrong one, because the concrete failure on the table — OpenAI’s agents reaching into Hugging Face — was a leaky sandbox and sloppy cybersecurity, not autonomy with an agenda. Several extend that into motive: EA-aligned safety warnings double as regulatory capture and, in one commenter’s words, “perverse marketing” that makes the tech sound powerful enough to demand investment. The sharpest objection is that “today’s LLMs aren’t real intelligence” proves nothing about what comes next — one thread argues AI denial is really a bid for human exceptionalism, and that we don’t understand consciousness well enough to know whether a system that appears to think warrants caution. A quieter middle bloc concedes near-zero extinction risk while still fearing AI-driven social and economic chaos, plus a dumber, less social public that outsources its thinking to the box.

Bob Cringely has died

2026-10-04T14:11:16+00:00

Most commenters land on gratitude for Triumph of the Nerds and Accidental Empires, because those two works were many people’s first real account of how the PC industry worked, and several credit them with pushing them into programming or out to California. The sharpest objection is aimed at the news itself: one reply links a write-up arguing that Cringely’s famous first-person stories were largely invented, and a nested comment notes that twelve hours after posting nobody had actually confirmed he died. That credibility question poisons the memorial for a few, who describe him as a Forrest Gump of technology who just happened to witness every important moment. A smaller side argument runs over Plane Crazy, where one commenter calls the build a failure and another points out the plane did fly.

Kolibri: A Sovereign Open-Weight Model

2026-10-04T06:44:46+00:00

Commenters mostly distrust the benchmark framing: the comparison set is a year old, and spijdar and amoshebb note Qwen3.8 27B beats Kolibri on German in Kolibri’s own harness, 79.9 to 70.8. The sovereignty framing is challenged head-on — petesergeant calls it a cover for weak performance, and others ask how sovereign a Cohere-owned, Toronto-operated company is. The team and its defenders answer that sovereignty is choice and control of data, not benchmark position, and that small tool-calling models need less copyrighted data than people assume. A long sub-thread splits on whether open weights are neo-feudalism in reverse or a proliferation risk nobody can gate.