Today’s thread is trust and control in AI systems: a major exposure at tl;dv left meeting recordings and transcripts accessible across tenants, while Illinois’ new age-signal law pushed debate about operating systems, Linux, and the spread of verification mandates. The same control questions run through AI tooling, from Docker’s sandboxed agent workflow and Claude Code’s new default auto mode to Meta’s local-agent model release and Mistral’s tool-calling patent filing. Elsewhere, security culture gets a reckoning in criticism of HackerOne, Anthropic’s math result prompts arguments about what counts as progress, and Squeak 6.1 brings a quieter programming-language release.
-
Tl;dv: Over 180k meetings left wide open
-
bobdahacker.com
- 173 comments
A security researcher says AI meeting startup tl;dv left more than 180,000 customer meeting recordings and transcripts accessible across tenants for months, including government and enterprise calls. The post also claims repeated private warnings to the CEO went nowhere for six months before public disclosure, which is why the story is drawing so much anger.
-
-
Illinois Just Passed a Law That Puts Linux on the Hook for Age Verification
-
linuxstans.com
- 343 comments
A new Illinois law would require internet-capable operating systems to carry a self-declared age signal by 2028 so apps and sites can treat minors differently. The Hacker News debate centered on whether this is a modest parental-control standard or the first legal foothold for broader age-verification and surveillance mandates that will spill into Linux and other open systems.
-
-
Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows
-
research.meta.ai
- 562 comments
Meta released Muse Glimmer, a 30B-parameter open-weight model aimed at always-on local agents on a single consumer GPU or high-RAM Mac/PC. The comments treated it as a serious new option in the 24-32GB VRAM class, with the real questions being token efficiency, tool use, context behavior, and whether it actually beats Qwen 3.6 in practice.
-
-
Docker Sandboxes – Disposable, isolated sandboxes for AI agents
-
docker.com
- 349 comments
Docker launched Docker Sandboxes, a local tool for running AI coding agents inside disposable microVMs instead of plain containers. The pitch landed, but the reaction centered on two practical knocks: it requires a Docker login for a local workflow, and many developers already have homegrown or open source alternatives with similar isolation, egress controls, and secret proxying.
-
-
What Happened to HackerOne?
-
blog.teknogeek.io
- 193 comments
A former HackerOne insider posted a long critique of how the bug bounty platform drifted from its hacker-first roots toward sales, AI products, and lower-quality handling of vulnerability reports. Commenters largely agreed the company has become more corporate and less trusted, but they also argued the underlying platform still solves a hard operational problem: paying and triaging security researchers around the world.
-
-
Auto mode is now the default in Claude Code
-
claude.com
- 302 comments
Anthropic says Claude Code will now default to “auto mode,” where a built-in model approves most commands instead of asking the user each time. The reaction was less about Anthropic’s benchmark numbers and more about what people actually do in practice: many already skip prompts entirely, but only after putting the agent inside containers, VMs, separate users, or other hard sandboxing.
-
-
Learning more about Claude's mathematical capabilities
-
anthropic.com
- 111 comments
Anthropic posted a research note saying an unreleased Claude model found a new result in analytic number theory: it raised a long-standing lower bound related to zeros of the Riemann zeta function from 41.6% to 67.2%, with papers, Lean formalization, and transcripts attached. Readers were struck less by the headline than by the workflow, where a non-mathematician mainly kept the model going with prompts like “believe in yourself,” and then argued over whether this is real mathematical progress, smart search over known literature, or marketing theater.
-
-
Mark Zuckerberg attacks 'closed' AI rivals as Meta returns to open models
-
ft.com
- 359 comments
A Financial Times piece and Zuckerberg essay argue that Meta is returning to releasing AI models more openly and attacking rivals that keep their best systems closed. The comments mostly agree open models are useful, but treat Meta’s stance as naked strategy from a company that lost the lead and now wants to commoditize the model layer.
-
-
Squeak 6.1
-
squeak.org
- 110 comments
Squeak 6.1 is a new release of the long-running Smalltalk system, a live programming environment where code, tools, and app state all live inside a running image. The comments turned it into a defense of image-based development and Smalltalk’s object model, with a lot more practical detail on where that approach still shines and where Squeak still feels dated.
-
-
Mistral Patent for “Code implemented tool calls”
-
patentsgazette.uspto.gov
- 170 comments
A US patent application from Mistral claims a broad method for letting an AI model generate code to call tools, run it in a sandbox, and continue from the results. Commenters overwhelmingly saw it as an obvious software pattern with lots of prior art, though a few noted the interesting part may be narrower claims around a resumable sandbox design.
-
-
Exploiting System Management Mode with a very long interrupt
-
github.com
- 39 comments
A GitHub writeup shows how a root-level attacker can abuse x86 System Management Mode by making one CPU core spend an extremely long time in a single instruction, letting firmware time out while waiting for all cores to enter SMM. The interesting part is not privilege escalation but a break in a hardware safety assumption, which raised fresh criticism of opaque firmware control paths like SMM.
-
-
Kinney Drugs pulls back AI phone assistant after hundreds of customer complaints
-
wcax.com
- 150 comments
A Vermont pharmacy chain rolled back a new AI phone assistant after hundreds of complaints about missed notifications and bad answers. The reaction centers on a simple point: customer-service AI is tolerable for narrow, low-stakes tasks, but pharmacy calls are full of messy exceptions where a fast handoff to a human matters more than clever conversation.
-
-
Japanese court overturns Red RAW video patent
-
dpreview.com
- 26 comments
A Japanese court invalidated a key Red Digital Cinema patent tied to recording compressed RAW video inside cameras. The patent had long been blamed for blocking camera makers from offering internal RAW recording without paying Red, so commenters treated this as a late but meaningful break in a widely disliked choke point.
-
-
Run Android ARM64 VR APKs on Apple Vision Pro
-
- 57 comments
A GitHub project called Klepton shows how to run Android ARM64 VR apps, packaged as APKs, on Apple Vision Pro. The post caught attention both as a technical hack around Apple’s locked-down headset platform and as a sign that the device may need unofficial tinkering to become broadly useful.
-
-
Humanising LLM Outputs Is Dumb
-
kuber.studio
- 84 comments
A blog post argues that making coding-agent output sound more human often makes it worse for both people and other agents, because style instructions change the model’s actual reasoning rather than just polishing the final wording. Commenters largely agreed that frontier models now produce bloated, jargon-heavy prose that hurts comprehension, though some said the real fix is better interfaces and separate rendering layers, not fighting the model in-prompt.
-
-
Tail-call optimization in C is relatively recent (2025)
-
lwn.net
- 111 comments
An LWN article traced how tail-call optimization in C compilers only became broadly usable fairly late, despite GCC having some support since 2001. The comments sharpened that into a more specific point: C still does not guarantee proper tail calls, but modern GCC and Clang now provide enough support, including `musttail`, that you can sometimes rely on them deliberately rather than hoping the optimizer helps.
-
-
How We Pushed CDC into Postgres
-
snowflake.com
- 28 comments
Snowflake published an engineering post on how it added change data capture, or CDC, directly into Postgres so customer database changes can be mirrored into Snowflake with less external infrastructure. The comments mostly treated it as part of the broader race to make OLTP-to-warehouse sync feel built-in, while arguing that the hard parts are still reliability, schema changes, and how much of the stack stays closed.
-
-
The Tragedy of the Cognitive Commons
-
arxiv.org
- 66 comments
A new paper argues that widespread AI use can deplete the shared social processes that create expertise, like apprenticeship, practice, and peer review, calling this a “cognitive commons” problem. Commenters mostly agreed the risk is real, especially for junior development and debugging instincts, but several pushed back that AI may simply shift skill toward higher-level judgment and expand who can build useful things.
-
-
Magnitude 7.4 Earthquake – 5 km S of San José del Palmar, Colombia
-
earthquake.usgs.gov
- 61 comments
A USGS page for a magnitude 7.4 earthquake near San José del Palmar, Colombia prompted firsthand reports from people in Medellín, Bogotá, and Pereira about long shaking, emergency alerts, evacuations, overloaded communications, and early signs of serious damage. The useful signal was less geology trivia than what actually held up in the first hour: phone alerts gave only seconds, normal telecoms bogged down, and satellite internet helped some families reconnect.
-
-
'Pervert glasses': Backlash against Meta's smart glasses grows
-
seattletimes.com
- 247 comments
A Seattle Times piece reports growing public backlash to Meta’s camera-equipped smart glasses, especially after users posted covert recordings of women and other bystanders online. The comments mostly agreed the issue is not just Meta’s reputation but the social damage of normalizing face-mounted recording that looks almost invisible to everyone else.
-
-
Sonic Pi v5
-
patreon.com
- 76 comments
Sonic Pi v5 is a new release of the free app for making and performing music by writing code, originally built for Raspberry Pi and now used more broadly on desktop systems. The post highlights friendlier errors, accessibility work, pro-audio features, and new audio-engine changes, while the comments focus on real-world usability, licensing questions around the new engine, and whether it is a serious tool or mainly a playful way into music.
-
-
Taxi drivers rarely die of Alzheimer's
-
theconversation.com
- 272 comments
A Conversation article argues that taxi and ambulance drivers seem less likely to die of Alzheimer’s because constant spatial navigation may build “cognitive reserve.” The comments mostly pushed back on the leap from an occupational correlation to a brain-protection claim, even though the underlying study did adjust for age and other basics.
-
-
Show HN: Needle2: 14MB agentic LLM for phones, wearables, smart home and robots
-
cactuscompute.com
- 63 comments
Cactus posted Needle 2, a 14MB on-device language model tuned for tool calling and structured extraction on cheap phones, Raspberry Pis, wearables, and even some microcontrollers. The pitch is not general chat, but fast local command-to-action mapping with a confidence score so products can execute locally or fall back to a larger cloud model.
-
-
Parametron: 50s Japanese computer that uses neither transistors nor vacuum tubes
-
ethw.org
- 45 comments
A computing history post highlights the parametron, a 1950s Japanese logic element built from magnetic components instead of vacuum tubes or transistors, and notes its use in early NEC computers. The useful signal is not just that it existed, but why this odd branch briefly made sense: it was cheaper and more reliable than the alternatives in Japan at the time, yet still hit hard physical limits that made it a dead end.
-
-
Study links GLP-1 drugs to bigger jump in women's employment than a degree
-
finance.yahoo.com
- 171 comments
A Yahoo Finance piece highlights a Harvard and NBER working paper claiming women who start GLP-1 weight-loss drugs see a bigger increase in employment than women who earn a college degree. Commenters largely accepted that appearance bias is real, but pushed hard on whether the paper actually shows "pretty privilege" or instead reflects health, mobility, and study-design issues.
-
-
Ask HN: In your experience, what are sound conventions for e-ink UI development?
-
news.ycombinator.com
- 47 comments
An Ask HN post asks for practical advice on designing browser-based apps for black-and-white e-ink phones, with specific pain points around ghosting, slow refresh, pagination, and streaming LLM output. The comments converged on a clear rule set: treat e-ink like print or a terminal, not like a modern animated touch UI, and build around discrete updates, strong contrast, and predictable page-level interactions.
-
-
Because It's Not Fun Enough: why languages fail
-
bytecode.news
- 134 comments
A blog post argues that programming languages usually fail because they stop being enjoyable enough to learn and use, not just because of raw technical merit. Readers pushed back hard, saying ecosystems, vendor backing, platforms, and tooling explain far more of language adoption than any notion of "fun."
-
-
An Interesting Fourier Transform – 1/f Noise (2007)
-
dsprelated.com
- 24 comments
A 2007 DSP article revisits a strange fact about pink noise, also called 1/f noise: power that rises as frequency falls, and a suggestive link between that shape and the Fourier transform. The comments turn it into a practical guide to what 1/f noise actually does in electronics, why it is hard to simulate and average away, and where the article’s mathematical hunch likely overreaches.
-
-
To recruit teachers, school districts are building homes
-
nytimes.com
- 65 comments
A New York Times piece says school districts facing teacher shortages are building or offering employee housing so staff can live near schools in places where normal housing has become unaffordable. Commenters mostly treated this as a symptom of a broken housing market, not a clever education policy.
-
-
Launch HN: Stoa Markets (YC S26) – A Marketplace for GPUs and AI Servers
-
stoaexchange.com
- 39 comments
Stoa Markets, a YC S26 startup, launched on Hacker News as a B2B marketplace for buying and selling new and used GPUs and AI servers rather than renting compute. The pitch is that standardized RFQs, escrow-like settlement, and dealer vetting can make a messy broker-driven market more liquid and eventually produce real resale data for financing.
-