Cammarata argues polysemanticity in early models came from fixable architectural and tooling failures, not superposition, and that most neurons in models up to trillions of parameters are legible, though thresholds vary.
Cammarata's central claim is that modern neural networks, up to trillions of parameters, are far more interpretable than earlier models.
OpenCode 2.0 rebuilds the API, adds a JavaScript code mode and background subagents, and runs one shared server per machine; 1.0 plugins and the old server API break.
OpenCode v2 addresses many V1 annoyances with automatic plugin and skill reloading, backgroundable tasks, a built-in configuration editor, improved front-end, and a loop reportedly re-architected around Effect.
NHTSA is reviewing a petition alleging about 1,000 Tesla robotaxis speed in school zones; the filing is tied to four incidents and mentions no crashes, and a review does not confirm a defect.
One owner reports that FSD 14.3 does not respect school zones even outside aggressive modes, expecting a hard speed-limit cutoff for any system acting on their behalf.
Cybersecurity experts attribute recent AI-agent incidents to missing containment, monitoring and egress controls, while labs dispute responsibility and liability remains unresolved.
Rogue-agent incidents generate publicity and can justify rules that large labs can meet but open-weight competitors cannot. The incentive is to appear dangerous enough to invite regulation, then shape it.
A solo founder reports Grok Bot agents handled support, engineering, security, marketing, and shipped a feature overnight; setup, data lock-in, and oversight remain caveats.
A critical agent system built on a third-party platform could be shut down with no recourse; a self-hosted option is sought.
AlphaNIST applies AlphaZero-style reinforcement learning to simulated NIST assembly tasks, but the supplied discussion questions its name and notes the lack of real-hardware validation.
Commenters question what AlphaNIST is and criticize a name that suggests Alphabet/DeepMind or the NIST OCR dataset; an author says it is an homage to AlphaGo/AlphaZero.
OpenAI says it used its own models to analyze the device; commenters see both a security failure and a PR misstep, while some doubt the espionage framing.
Prompts to an LLM are recorded and can be used in prosecutions, echoing search-history evidence; one reply argues a MAC-address search on company property is not necessarily abuse.
Google claims Gemini 4 Argon advances robot reasoning and control, but the supplied discussion questions whether on-device power, cost and specialized models will limit general-purpose deployment.
A comment estimates that even with continued model efficiency and hardware power-per-watt gains, a battery-powered Argon model on a non-bus-sized robot may be about a decade away, with additional years needed for…
Anthropic's Scion shows agent-scale monorepos can outgrow git, but the supplied discussion questions the traffic figures and whether human review survives the shift.
Replies to the proposal to abandon source control argue that agents still need branches, pull requests, tags/versioning, build reproducibility, and history to understand changes and detect revert loops.
A HN post on a feed reader drew heavy AI scraper traffic, with one service accounting for nearly half of requests; the author responded with blocking and rate limits. The sample is small and cannot show broader patterns.
A commenter reports blocking major cloud ASNs (AWS, Google Cloud, Azure, Alibaba, Oracle) due to spam, vulnerability scanning, and aggressive scraping, and argues providers must control abuse under their networks.
Google's Gemini 4 Argon page claims broad engineering and scientific gains, but the only supplied comment criticizes the page's mobile usability, not the model.
Commenters question whether the pelican-on-bicycle SVG test still measures model capability, since it is a known meme likely absorbed into training data and Google has specifically trained for SVG output.
Character.ai's Chapter 11 filing is a court-supervised sale attempt after legal and operational strain; comments frame it as a shell left by a Google talent deal, with wind-down possible if no buyer emerges.
Comments describe Character.ai as effectively acquired after Google hired the founding team and licensed technology, leaving a Chapter 11 shell.
Apple's Minions paper shows idle Macs can serve as a task-parallel inference pool, but commenters doubt Apple will sell rackable hardware or build a broader cloud.
Comment 50071535 describes Minions as an internal Apple system that pools idle Mac GPUs into a shared compute cluster for ML inference.
Character.AI's Chapter 11 filing and proposed Disney asset sale cap a turbulent run, but the supplied discussion is a small sample split between doubts about value creation and concern over user data.
A commenter asks what will happen to the very personal and sensitive data the company likely holds on users. A reply suggests it could be auctioned off in bankruptcy, citing Spirit Airlines as an example.
EmbeddingGemma 2 is repurposed into a 288× on-device English transcriber, but its ~5% WER trails leading speech models and multilingual support remains unproven.
A commenter asks whether fast transcription requires playing audio at 200x speed.
