Google Search Is Dying. What Comes Next Is Worse

810 points · The Walrus

Vass Bednar’s piece in The Walrus argues that AI is cannibalizing the open web, and the internet’s collective memory is disappearing as a result. The hook: Google’s AI Overviews are now hallucinating sunset times — a query with a deterministic, publicly available answer — because the system is synthesizing responses from scraped content rather than pointing you to the source. Users planning outdoor events are getting confidently wrong answers from a system that was supposed to make search better.

The broader thesis is that Google’s AI summaries are hollowing out the incentive structure that made the web worth indexing in the first place. Publishers lose traffic, creators lose attribution, and users get a probabilistic approximation of knowledge instead of verified sources. The Walrus frames this as a Canadian media story — local publishers watching their referral traffic evaporate — but the dynamic is global. When the search engine that monopolized information discovery decides to keep users on its own pages, the downstream effects on journalism, independent blogs, and niche reference sites are existential.

The uncomfortable truth the article surfaces: we built the most comprehensive knowledge infrastructure in human history, then handed its gatekeeping to a company whose revenue model requires maximizing engagement on its own properties. The “next” in the title refers to AI-native search startups, but the real concern isn’t which interface you use — it’s whether the content being summarized will continue to exist at all.

Needle 2: 14MB Agentic LLM for Phones, Wearables, and Smart Home

495 points · Cactus Compute

Cactus Compute released Needle 2, an open 45-million-parameter model for tool calling, device control, and structured extraction. The entire model is a single 14MB binary running in 28MB of RAM. It’s built on their Simple Attention Network architecture, compressed to CQ2-bit with Cactus Quants, and ships with its own inference engine — no external runtime dependency.

On benchmarks, Needle 2 trades wins with models 5× to 70× larger (FunctionGemma 270M, LFM2.5 230M, Apple FM) while running at 2-bit precision against their fp16. Claimed throughput: 500 tokens/sec decode on Raspberry Pi 5, 400–1,500 on VR headsets (Meta Quest 3S, Apple Vision Pro), and 300–700 on sub-$200 phones. Peak session RAM is around 28MB, which puts it within reach of newer microcontrollers like the ESP32-S3.

The demo is impressive — multi-step tool routing, device control, form filling, and refusal of off-topic queries, all running as WebAssembly in the browser. The question is whether “tool calling” in a 14MB model means reliable structured output in production, or just passing controlled demos. The benchmarks they cite (Mobile-Actions, BFCL) are reasonable choices, but real-world agent loops are messier than eval suites. Still, the size-throughput tradeoff is genuinely novel. If you’re building on-device AI for embedded systems, this is worth a serious look — the alternative is shipping a 270MB+ model or giving up on local inference entirely.

England Set to Eliminate Hepatitis C

427 points · BBC

England is on track to become one of the first countries to eliminate hepatitis C as a public health threat. The 80% treatment target for known cases has already been met, and deaths from the virus have fallen 36% over the last decade — just short of the WHO’s 2030 elimination threshold. Modern antivirals cure more than 95% of cases in 8 to 12 weeks of tablet treatment.

The program’s success hinges on aggressive case-finding: testing in A&E departments, outreach to homeless populations and people who inject drugs, and opt-out screening in prisons. These are populations that traditional healthcare systems routinely miss. The BBC article notes that England is ahead of most comparable countries, which says less about England’s brilliance than about how poorly most nations have executed on a problem that is now technically solvable.

The 300+ comments suggest this resonated deeply. Hepatitis C elimination is one of those public health wins that’s invisible when it works — nobody celebrates the liver cancers that didn’t happen. The broader lesson: when you combine a genuinely effective treatment (direct-acting antivirals are a marvel) with systematic screening infrastructure, elimination is achievable. The bottleneck was never the science; it was the logistics and political will to reach the hardest-to-reach patients.

H3-Metal: Native MiniMax-H3 Inference for Apple Silicon

405 points · antirez/h3.c

Salvatore Sanfilippo — Redis’s creator, now apparently doing native video generation on Apple GPUs — released h3-metal, a native inference engine for MiniMax’s H3 model on Apple Silicon. The project generates video and audio from text prompts, with first/last-frame conditioning and ordered reference image support, all running through Metal directly on Mac hardware.

The build is structured as sequential vertical slices: deterministic metadata first, then Metal block parity, prompt encoding, prompt-to-video/audio, and frame conditioning. The current state has all of these working end-to-end, with ongoing optimization for M3 Max and M5 Max. An interactive CLI session lets you generate videos, set seeds, adjust frame counts, and chain reference images — a surprisingly polished developer experience for what’s essentially a solo project reverse-engineering a proprietary model’s inference path.

The significance here is less about MiniMax H3 specifically and more about the pattern: a world-class systems programmer spending his time making a Chinese AI company’s video model run natively on Apple hardware. This is what happens when the inference layer becomes commoditized — the interesting work moves to making it fast on specific silicon. The 42K lines of README suggest serious documentation discipline, too.

Stealing Reasoning Traces from Proprietary LLM APIs

348 points · Stolen Thoughts

A research team demonstrated that encrypted chain-of-thought reasoning traces from Anthropic, OpenAI, and Google can be recovered in plaintext through a two-step API attack. The method: take an encrypted reasoning trace from a frontier model (e.g., Claude Opus), replay it into a weaker sibling model from the same provider, then jailbreak the weaker model to recover the stronger model’s hidden reasoning — without ever directly attacking the stronger model or triggering its anti-distillation safeguards.

The researchers show that these encrypted reasoning blocks are replayable across sessions, users, and models. The attack exploits the fact that the “encryption” on reasoning traces is more obfuscation than cryptography — the traces can be fed as input to related models that share the same tokenization and reasoning format. From there, a weaker model can be prompted to “continue” or “explain” the reasoning, effectively decoding it.

This is a significant finding for anyone building competitive moats around reasoning capabilities. If your chain-of-thought can be extracted by buying API access to a cheaper model in your own lineup, the reasoning trace is not a trade secret — it’s a leaky abstraction. The paper names specific models from all three major providers. The practical implication: if you’re shipping encrypted CoT to clients, you’re shipping your reasoning to competitors. The HN discussion (133 comments) is predictably heated, with debate over whether this constitutes “stealing” or simply observing what the API voluntarily returns.

Chicken Scheme 6.0

291 points · code.call-cc.org

Chicken Scheme 6.0 is a major release that’s been a long time coming. The headline changes: internal string representation is now UTF-8 (full Unicode support), all R7RS-small modules are available in the core system, and bytevectors have replaced the old blob type with R7RS-compatible operations. This is a significant modernization — Chicken was one of the last mainstream Scheme implementations running on Latin-1 strings by default.

Other notable changes: process management now returns process objects instead of raw PIDs, file-locking uses flock(2) and is thread-safe, and port constructors take keyword arguments. The make-binary-input-port and make-binary-output-port additions are practical — binary I/O was awkward before. Locatives on strings are now indexed by code-point rather than byte, which is the correct behavior for a Unicode-aware system.

Chicken has always occupied an interesting niche: a Scheme that compiles to C, ships as a shared library, and plays nicely with the C ecosystem. The 6.0 release brings it up to modern language baseline expectations (Unicode, R7RS compliance) without breaking the compilation model that makes it useful for embedding. For the 291 people who upvoted this, it’s probably less about any single feature and more about a beloved niche language demonstrating it’s still alive and shipping meaningful releases.

Apple Silicon VMs: 11–16× Faster LLM Inference with llama.cpp

254 points · trycua/cua

The Cua team (macOS virtualization stack) discovered that Apple’s Virtualization.framework reports conservative GPU capability profiles to guest VMs — specifically, an Apple family 5-era profile with 32KB max threadgroup memory and no SIMD-group matrix support. This forces llama.cpp to take slower Metal code paths even though the host GPU supports the newer ones. Their fix: a process-scoped Metal capability shim that intercepts capability queries and returns accurate values.

The results are dramatic. On an M1 Ultra with TinyLlama 1.1B: prompt processing went from 431 tok/s (stock VM) to 4,786 tok/s (unlocked) — an 11× speedup that’s 98% of bare-metal performance. Token generation hit 16× faster. With Gemma 4 12B QAT, the unlocked VM reached 99.59% of bare-metal prompt speed. Even Meta’s Muse Glimmer 30B showed 7.5× and 8.8× improvements for prompt processing and generation respectively.

The clever part is that this isn’t actual GPU passthrough — it’s still running through Apple’s paravirtualized virtual GPU path. The shim just tells llama.cpp the truth about what Metal features the hardware actually supports, so it selects optimal kernels instead of conservative fallbacks. This is the kind of bug that likely affects every Virtualization.framework user running Metal workloads, not just LLM inference. The source is released so others can validate and extend the approach.

Nvidia’s Risky Business

232 points · Stratechery

Ben Thompson opens with Jay Cooke and the Northern Pacific Railway — a financing scheme where generous commissions drove increasingly risky bond sales to retail investors, eventually contributing to the Panic of 1873. The parallel to Nvidia’s current position is the throughline: a company whose success has created incentives that may be pushing it into structurally risky territory.

The Stratechery argument (behind a paywall, so this is from the freely available portion) positions Nvidia at an inflection point where its dominance in AI accelerators is being challenged not by better chips, but by customers designing their own (Google TPUs, Amazon Trainium, Microsoft Maia) and by the inference market fragmenting into edge devices, phones, and specialized hardware. Nvidia’s response — expanding into software platforms, networking (Mellanox), and cloud services — mirrors the vertical integration moves that historically signal a hardware company worried about commoditization.

The 232 points suggest moderate interest, likely split between people who think Thompson is overstating Nvidia’s risk (their revenue growth is still extraordinary) and those who recognize the pattern: when your best customers become your competitors, the margins eventually follow. The Cooke analogy is apt — not because Nvidia will collapse, but because the incentive structure of selling AI infrastructure at current margins is attracting exactly the kind of competition that will compress those margins over time.

The Throughline

Today’s HN front page tells a story about infrastructure reaching its limits and being rebuilt at different scales. The Walrus piece shows information infrastructure (Google Search, the open web) being hollowed out by the very AI systems it enabled. Needle 2 and h3-metal represent the counter-movement: pushing inference infrastructure down to the smallest possible hardware and out to Apple Silicon, respectively. The Cua VM findings reveal that even Apple’s own virtualization infrastructure is silently bottlenecking the workloads people care most about.

The stolen reasoning traces paper cuts deepest: it exposes the assumption that encrypted chain-of-thought is a competitive moat. If your reasoning can be replayed into a cheaper model and decoded, then the entire “reasoning trace” feature that Anthropic, OpenAI, and Google are shipping is a security theater masquerading as IP protection. This has immediate implications for anyone building on these APIs — your prompt engineering secrets may be extractable by anyone with API access to a related model.

Chicken Scheme 6.0 and England’s hepatitis C elimination are the quiet wins — a niche language modernizing its foundations and a public health system actually executing on a solvable problem. Both are reminders that boring, sustained work on infrastructure (whether Unicode support or opt-out prison screening) compounds in ways that flashy launches never will. Meanwhile, Thompson’s Nvidia analysis asks whether the company building the picks-and-shovels for the AI gold rush is itself overextending. The answer probably depends on whether you think AI infrastructure demand is a railroad bubble or a genuine paradigm shift — and history suggests it’s always a bit of both.