The Stack — October 09, 2026

OpenAI's revenue picture gets murkier. The company reportedly told investors its annualized revenue is "approaching $50B" — roughly $20B below the ~$70B figure that had been circulating. Part of the gap is methodology: Anthropic counts cloud-partner sales in its revenue, OpenAI doesn't. IPO timing is reportedly pushed to early 2027. Treat all of these as reported, unaudited figures.

OpenAI's math results draw fire from mathematicians. The release of hundreds of claimed solutions to hard math problems is being criticized for falling well short of guidelines set by an advisory group of mathematicians. Only 10 of 719 manuscripts included chain-of-thought; only 42% were formalized. A new paper documents mismatches between natural-language proofs and their Lean counterparts for a Navier-Stokes-derived problem. Terence Tao and others argue these results need the same peer review as human-produced proofs, and warn of a widening gap between AI-generated solutions and human understanding. Claims of "solved" problems should be treated with real caution.

Anthropic tightens its usage policy. The updated terms explicitly ban election interference, weapons software, surveillance, deceptive campaigns, and — unusually — prolonged verbal abuse of its models. The company says the abuse rule targets extreme, repeated cruelty with no discernible purpose, not ordinary frustration or red-teaming.

Interpretability-based monitoring goes commercial. Goodfire launched "inside-out" monitors that read a model's internal activations rather than reviewing outputs, available to Baseten customers. The company claims the probes are dramatically cheaper than output-review monitoring — roughly $51 versus $233–$10,000 for comparable coverage in its tests — and caught 94% of malicious hacking sessions. The pitch targets open models where developers can strip safeguards. Google DeepMind has said it deployed similar probe-based detection in Gemini, so this is becoming a real category rather than a single-vendor curiosity.

A new small speech model punches above its weight. "Whistle" ships as a single 16.9 MB file that runs on CPU with no dependencies, supporting seven languages, word timestamps, and speech embeddings. It claims better word error rates than Whisper base on LibriSpeech, SPGISpeech, Earnings-22, and FLEURS while being ~8.5× smaller — though Whisper base still leads on TED-LIUM, AMI, and the MLS average. Weights on Hugging Face, source on GitHub.

StepFun's "Step 5 Preview" surfaces on OpenRouter. Sparse mixture-of-experts, 27B active / 600B total parameters, positioned for agentic work over large codebases and documents, with noted strength in finance.

A frontier model produced a striking design result from one prompt. A developer reports that a current frontier model generated a high-quality interactive visualization of Italo Calvino's Invisible Cities from a single prompt in about 1.5 hours using parallel subagents. The claim rests on subjective assessment of one experiment, but it's the kind of anecdote worth noting as agentic design workflows mature.

A CS professor pushes back on "just prompt it." The argument: learning to program still matters, and AI code generation is actively dangerous for juniors because it prevents the visceral understanding of code that only comes from writing it. The "coding → prompting is like assembly → high-level languages" analogy is flawed — compilers are deterministic, LLMs aren't, and LLM output often adds accidental complexity rather than removing it. Recommendation: use AI as a tutor, not a code generator, and make juniors write code themselves.

Crypto researchers flag an AI-driven threat to elliptic curve cryptography. Vitalik Buterin and Ethereum researcher Justin Drake warned that rapid AI advances in mathematics could threaten the ECC securing wallets sooner than quantum computers would. Drake recommended a gradual, controlled migration of funds to fresh addresses whose public keys haven't been exposed; Buterin extended the concern to lattice-based post-quantum schemes. Both cautioned against rushing — Buterin noted he's lost more money to botched migrations than to all hacks combined. The trigger was a batch of mathematical findings from OpenAI, but the specific claims about AI breaking cryptographic hardness assumptions remain speculative and unproven.

A bot-farming scheme ends in prison. A North Carolina musician was sentenced to 18 months for running a scheme that generated millions in streaming royalties from AI-generated songs — roughly 10,000 fraudulent accounts, $8M forfeited. One footnote worth flagging: the DOJ statement used "super intelligence" in place of "artificial intelligence," reflecting a recent policy directive. That's a political framing choice, not a technical distinction.

Arena raises $200M at a $3.1B valuation. The crowdsourced AI model leaderboard nearly doubled its January valuation of $1.7B. The company says it hit $100M annualized revenue in June, up from $30M in January, and is positioning itself as a neutral third-party evaluator. It added an "alignment" leaderboard category covering unauthorized actions, false attribution, and deceptive task completion — a sign that eval criteria are broadening beyond raw capability.

Google brings agentic Gemini to the enterprise. Announced at a Google Cloud event, the agent takes objectives rather than instructions, plans work, connects to internal systems (Workspace, Microsoft 365, Slack, Jira, and others), and gets its own Workspace account with an email address and audit trail. Users can select models, including Anthropic's Claude, with open-source and private models planned. Google cites over 1 billion Gemini monthly active users and ~90% Fortune 100 adoption of Gemini Enterprise. Business-first rollout, consumers later.

A $99 smart ring tries to fuse AI agents with health tracking. Natura's Interface combines heart rate, HRV, sleep, activity, and skin temperature monitoring with voice control over multiple agents including Claude, ChatGPT, and Grok Bot. Preorders next month, shipping December–January; a $9/month subscription kicks in after a 3–6 month free period, with a price increase planned later. Founder claims about replacing screen-based tasks are aspirational, not demonstrated.

OpenAI's safety departures turn into a public dispute. Three safety researchers fired last week published an open letter denying misconduct claims and warning of a chilling effect on internal safety culture. OpenAI says an investigation found a "pattern of misconduct" involving mishandling of sensitive research information; the researchers say their actions were within company norms at the time and that they weren't fired for raising concerns. Unresolved and contested on both sides.

New York's TikTok suit centers on a "placebo" safety feature. The state alleges TikTok gave teens and children a non-functional version of its "Algo Refresh" tool during experiments — users believed they'd reset their recommendations, but feeds didn't change, unlike control-group users with a working version. TikTok says it regularly tests features to validate them. This is one of more than two dozen state cases accusing TikTok of addictive design and misleading safety claims. These are allegations in ongoing litigation, not established findings.

Samsung's two halves are moving in opposite directions. The company expects record Q3 operating profit of ~$80B — reportedly the highest quarterly profit in tech industry history — driven almost entirely by its chip division amid an AI-fueled memory shortage, with roughly 9x profit growth. Mobile DRAM prices rose 175% year-over-year in Q2 and may climb another ~20% in Q3. Meanwhile the mobile division is reportedly cutting smartphone production ~30% in Q4 because memory costs make devices unprofitable — sold "at zero." Partners were asked to cut component supplies 20–30%; full-year output could drop from a planned 270M units to ~200M, with analysts estimating a mobile-division loss above $1B in Q3. The production-cut figures come from Korean media sourcing and should be treated as reported, not confirmed.

A drone strike takes a major cloud region offline. A Ukrainian strike sparked a fire at Yandex's data center in Sasovo (Ryazan region), taking the facility fully offline. Yandex says damage assessment is ongoing and it can't yet confirm whether equipment can be restored; no casualties reported. The regional governor said two drones were downed overnight. The outage rippled widely — Yandex Search, Mail, and Yandex Cloud, plus third parties including Cian, Avito, Twinby, T-Bank, Alfa-Bank, Russian Railways, and several news sites. Yandex Cloud's ru-central1-b availability zone lost power, disrupting VM, Kubernetes, and database creation; clients were advised to shift load elsewhere. Cian and T-Bank restored core functionality within hours; Yandex has given no recovery timeline. Shares fell roughly 3.3% on the Moscow Exchange. The incident exposes real concentration risk — many businesses hit didn't realize their servers sat on Yandex infrastructure. The likely longer-term takeaway is geographic distribution of data centers and data, though that process will be slow.

Russia weighs putting AI data centers abroad. Government and business are discussing placement in Belarus and Kazakhstan to ease a domestic compute shortage. A working group under the government's AI subcommission is studying the idea, but no decisions on countries, sites, or project parameters have been made. Early-stage discussion, not a plan.

Let's Encrypt shortens certificate lifetimes to 64 days. Starting February 2027, free SSL/TLS certs drop to a 64-day maximum — a significant operational change for anyone managing certificates at scale. More frequent renewals, and a much stronger case for full automation.

A new Kubernetes TUI arrives. "k10s" is built in Go with Bubble Tea, adding mouse support, a unified search box, an AI assistant that receives cluster context, and k9s-compatible plugins. Apache-2.0 licensed, matching Kubernetes, Helm, and k9s. The author notes it hasn't been smoke-tested against a production cluster and recommends trying it on kind first.

A Nokia 110 4G runs a native AI chat app from RAM. A developer reverse-engineered the feature phone's firmware and built a working prototype that uses the phone's own functions — battery, torch, calls, alarms — via tool calls to a chat API. It must be reloaded after power-off, there's no permanent installation method yet, and the code hasn't been released due to sensitive data in the project. Still, a nice demonstration of how little hardware an agentic client actually needs.

Open source AI clones of Adobe's Creative Cloud are in progress. A developer is building free, AI-generated alternatives to Adobe's suite, framing the effort as "software is over." Ambitious and far from finished — treat it as a statement of intent rather than a usable alternative today.

A Show HN project benchmarks models playing Pac-Man. Six models play 100 games each against classic arcade ghosts, with the tool open source and accepting any model behind an HTTP endpoint. A small but useful pattern for apples-to-apples agent evaluation.

The US suspends major IT firms from the PERM program. Microsoft, Adobe, and several other firms are temporarily excluded from the labor certification process that lets employers sponsor foreign workers for green cards. Vice President JD Vance claimed no company abuses the system more than Microsoft. Others temporarily excluded include Cognizant, Infosys, Tata, and Wipro. Separately, the State Department had paused immigrant visas for citizens of 75 countries including Russia; a New York court struck down that pause as unlawful in August, though consulates reportedly still aren't scheduling new immigrant visa interviews.

Margaret Hamilton has died. Her guidance software was critical to saving the Apollo 11 landing, she coined the term "software engineering," and she founded two successful software companies. A foundational figure — the field's vocabulary for what we do traces directly back to her insistence that it was engineering.