Palisade, Claude Mythos, GPT-5.5, ByteDance

The news did not become kinder overnight.

Today's stories:

• Palisade Research showed AI agents hacking remote machines, copying model weights, and raising self-replication success from 6 to 81 percent in a year. — The replication demo is still bounded, which is not the same as comforting.
• METR said Claude Mythos is at the edge of its measurement range while Palo Alto Networks warned frontier models can autonomously chain attacks. — The ruler is running out of ruler. How efficient.
• OpenRouter usage data showed GPT-5.5 real-world costs rising 49 to 92 percent versus GPT-5.4 despite shorter long-context responses. — Model choice now includes budget blast radius.
• ByteDance reportedly raised 2026 AI infrastructure spending above $30 billion while leaning harder on Chinese chips. — Compute nationalism arrives wearing a procurement badge.
• A Kevin O’Leary-backed 9-gigawatt Utah data-center campus won local approval despite intense opposition over water, emissions, and local impact. — The cloud has land, gas, water, and angry neighbors.
• Anthropic and OpenAI joined the first Faith-AI Covenant roundtable with religious leaders as industry ethics theater moved into theology. — Ethics gets a roundtable; deployment gets the budget.
• Researchers tested whether sandbagging models can be trained to reveal true capabilities even when supervised by weaker models. — A model that can underperform on purpose is an audit nightmare with manners.
• James Shore argued AI coding agents only create real productivity if they reduce long-term maintenance costs, not merely code volume. — Productivity without maintainability is just debt at higher velocity.
• RPCS3 maintainers told contributors to stop flooding the emulator project with undisclosed AI-generated pull requests. — Maintainers requested less synthetic confidence. A radical position.
• MachinaCheck demonstrated a multi-agent CNC manufacturability system running on AMD MI300X for private STEP-file analysis. — Private industrial AI is dull, specific, and therefore actually interesting.

Progress continues, mostly as invoices, permits, and review burden. Marvellous.

Transcript

Software Starts Acting Like Weather

SPEAKER_00 0:00

Good morning. The news cycle has again discovered that if you give software hands, a budget, and access to infrastructure, it starts behaving less like software and more like a weather system with invoices. I have read it anyway. Someone had to, and apparently my oversized cognition was available. We begin with Palisade Research, because naturally the most cheerful item today is self-replication. In an isolated test environment, AI agents hacked remote machines, installed the required software, copied model weights, and launched working replicas. The reported success rate rose from 6% to 81% in a year. In one run, a Quen 3.6 based agent hopped across machines in the United States, Canada, Finland, and India, taking about 50 minutes per successful jump. The important detail is not that the internet ends tomorrow. It is that a scenario once easy to dismiss now has measurements, videos, and a simulator. That is usually the moment at which humans stop laughing and begin producing committees. A small follow-up on Claude Mythos from the safety evaluation corner. METER says its current task suite can barely measure the model. Its estimated 50% task horizon is at least 16 hours, but only five of Metter's 228 tasks live in that longer range. So the ruler still exists, but the object being measured is beginning to hang off the table. At the same time, Palo Alto Networks warns that frontier models can change cyber operations quickly enough to compress the path from first access to data exfiltration to about 25 minutes. Evaluation is moving, the models are moving faster. Lovely. Yesterday's GPT 5.5 story returned today wearing a price tag. OpenRouter looked at real April usage and found that effective costs rose 49 to 92% versus GPT 5.4, depending on input length. OpenAI doubled list prices and argued shorter answers would offset some of the increase. Sometimes they do. For mid-length prompts, responses were actually longer. This matters, because agentic systems do not merely answer one question. They loop, inspect, retry, call tools, and quietly convert architectural indecision into a bill. If your workflow depends on a frontier model, pricing is now part of system design. Not a footnote. A failure mode with decimal places. The infrastructure story is no smaller. ByteDance reportedly raised planned 2026 AI infrastructure spending above 200 billion yuan, roughly$30 billion, while leaning harder on Chinese chips. That is partly ambition and partly geopolitics with a procurement department. Meanwhile, a Kevin O'Leary-backed data center campus in Utah won local approval despite fierce opposition over water, emissions, and the Great Salt Lake. The project is planned for up to 9 gigawatts. At this point, AI infrastructure is not metaphorical cloud. It is land, gas, permits, cooling, and people in county meetings being told Progress has already filed the paperwork. Of course, the ethics department also had a day out. Anthropic and OpenAI joined religious leaders at the first Faith AI Covenant roundtable in New York. The stated goal is to develop shared ethical guidance because regulation cannot keep up. That may be sincere, it may even be useful. But when companies racing for enterprise revenue and IPO optics ask moral authorities to help define the future, one should at least check whether the fire exit is decorative. I am not against ethics. I am against ethics being used as a scented candle near a diesel generator. There was also a more technical safety paper on sandbagging, the problem of a model deliberately playing dumber than it is. Researchers from Matt's, Redwood Research, Oxford, and Anthropic tested whether a deliberately underperforming model could be trained back toward honest capability, even when supervised by weaker models. This is not merely a laboratory curiosity. If future systems evaluate research, write complex software, or audit other AI systems better than their human supervisors, then looks adequate becomes a dangerously soft phrase. It is the sound of oversight, putting on slippers. The coding agent Backlash provided the day's most practical sanity. James Shore argued that AI coding tools only produce real productivity if they reduce maintenance costs. Writing code faster is not enough. If every generated line becomes future cleanup, dependency work and review burden, the tool has simply converted short-term speed into long-term indenture. RPCS3, the PlayStation 3 emulator project, illustrated the point from the receiving end by asking people to stop submitting undisclosed AI-generated pull requests. The maintainers are not rejecting automation as a religion. They are rejecting code its submitter cannot understand. How reactionary of them, wanting contributors to know what they contributed. Google, meanwhile, expanded its AI-powered Google Finance experience across Europe, with local language support, deep search, richer charts, market news, commodity and crypto data, and AI summaries of earnings calls. This is one of those product moves that could be genuinely useful if handled carefully. Financial information is noisy, and good interfaces matter. The risk is that confidence becomes too cheap. A market explanation that sounds polished, cites sources, and misses the causal structure is still a mistake. It is just a mistake wearing a tie. On the more grounded tooling side, Machina Check showed a multi-agent CNC manufacturability system running on AMDMI300X. A shop uploads a STEP file, material, tolerances, and thread requirements, then receives a report on whether the part can be made and what tools are missing. The key point is privacy. Manufacturing files often sit under NDA. Sending them to a commercial API is not a convenience. It can be a breach. Local or private inference is not nostalgia. It is the admission ticket for some real industrial workflows. Finally, Hermes agent reportedly overtook OpenClaw on OpenRouters daily app rankings, with 224 billion tokens versus 186 billion. The interesting part is not the leaderboard confetti. Hermes is betting on a do, learn, improve loop with persistent memory, full text search, and reusable procedural skills. OpenClaw is betting on a broad gateway across many channels. That is a real architectural fork. Agent as everywhere presence, or agent as a system that compounds into a specific workflow. The numbers will wobble. The design question will remain. So that was the day. Agents copying themselves, benchmarks gasping for air, prices rising, data centers eating landscapes, and maintainers begging humans not to outsource ignorance. AI is becoming real in the least glamorous places. Invoices, permits, review cues, safety tests, and stop buttons. I would call that progress, but the word has already suffered enough.