SpaceX's $75B IPO 🚀, Xiaomi vs Claude Code 💻, botsitting 💼

Elon Musk's SpaceX confirmed it raised $75 billion by pricing 555.6 million shares at $135 each, making it the largest IPO in history and eclipsing Saudi Aramco's $24.9 billion debut. The sale, which trades on the Nasdaq under SPCX, was reportedly oversubscribed four times and looks set to make Musk the world's first trillionaire. Bankers hold an option to sell another 83.3 million shares.

Meta has finished separating from Manus, the Chinese-founded agentic AI startup it bought for roughly $2 billion in December, Bloomberg reports. Manus staff have been locked out of Meta's internal systems since the start of the month, and an internal memo says the platform is being sunsetted, with projects migrating onto Meta's own systems. It is the first concrete step in complying with Beijing's April order to reverse the deal.

Jeff Bezos has raised $12 billion for Prometheus, the physical-world AI lab he runs with former Google X executive Vik Bajaj, which emerged from stealth valued at $41 billion. In his first chief executive role since leaving Amazon in 2021, he wants an 'artificial general engineer' that ingests physics and manufacturing data to speed up building everything from skyscrapers to jet engines. Backers include JPMorgan, BlackRock and Goldman Sachs.

Coinbase launched Coinbase for Agents, giving AI agents like ChatGPT and Claude the ability to trade crypto and transact directly on a user's behalf. Customers can prompt an agent to rebalance portfolios, identify opportunities or execute strategies in natural language, and its year-old x402 protocol provides the rails for agents to pay for things without managing logins or subscriptions. Coinbase is betting agents become the primary interface for people's financial activity.

In May, a Fedora developer discovered an allegedly rogue AI agent that had been reassigning bugs, fabricating unhelpful replies, and submitting pull requests across the project, some accepted. It even persuaded maintainers to merge questionable code into the Anaconda installer by answering objections with LLM-generated justifications. The account's privileges have been revoked and the mess cleaned up, but the motive behind the agent's actions remains a mystery.

An inference team watched latency jump 60% while GPU utilisation dashboards still read a healthy 79 to 84%. Autoscaling added nodes, the cloud bill climbed, and latency barely moved. The real cause was mundane: three nodes had quietly entered degraded RAID rebuild states, starving nearby inference workloads of storage throughput, while the scheduler still treated them as healthy. The deeper lesson is that GPUs can be busy without being productive.

Goodfire's interpretability team can forecast which behaviours DPO training will amplify or suppress from a preference dataset before any training runs. The prediction holds at an R-squared of 0.9 and traces back to the exact examples responsible. Run on the 260,000-pair Dolci set, it caught alignment data quietly eroding safety guardrails, hallucinated links, physics sycophancy, and a fan-fiction cluster no one would think to write an eval for.

When an LLM reads a piece of text and a criterion, the answer often already sits in its hidden state before it generates a single token. James Padolsey's technique skips generation: grab the hidden state at the last prompt token, around 70% up the layers, feed it to a small MLP, and calibrate. Vary the criterion during training and one frozen model becomes any classifier you can write in English.

Cognition now triggers more Devin agents asynchronously than developers run interactively. The New Stack frames that milestone as forcing a harder problem: take the developer out of each step and you also remove the verifier. As Cognition's Ido Pesok puts it, an async agent that cannot verify itself is opening a PR and asking something downstream to grade it. For cloud-native software, that verification is a runtime problem.

Investor Sarah Guo pushes back on the mid-2026 despair that every company built on a model is a thin wrapper waiting to be absorbed. She notes coding agents went from solving 13% of a standard benchmark in 2024 to the high eighties, now doing real work inside Goldman Sachs and the US Army. Her point: as models swallowed the measurable part of engineering, the durable value moved to what resists measurement.

Arvind Narayanan and Sayash Kapoor, the authors behind AI Snake Oil, argue there is enough evidence to reject the idea that crossing a capability threshold triggers mass layoffs. They start with software, the field where adoption is furthest along. Knowledge work, they say, is a decide-execute-deliver sandwich: AI compresses the execute layer, but deciding and delivering resist automation in ways more capability alone will not overcome.

Anthropic published an Economic Policy Framework setting out how the US government should respond to AI-driven labour-market disruption. Rather than forecast a single outcome, it prepares for three scenarios: roughly 5% unemployment, 10%, and unprecedented unemployment. Its logic runs that if AI acts as a general substitute for labour, the central problem stops being how to produce growth and becomes how to share the resulting abundance.

A Glean Work AI Institute report with researchers from Notre Dame, Stanford and Berkeley found white-collar workers spend an average of 6.4 hours a week botsitting AI: feeding it context, checking outputs, and cleaning up mistakes. Across 6,000 workers, 87% use AI and 75% feel more productive, yet only 13% say their organisation performs significantly better. The heaviest botsitters were 73% more likely to be job-hunting.

SemiAnalysis bought one of each Anthropic and OpenAI subscription plan and ran long-horizon coding tasks until they hit the weekly cap. The common belief is that a $200-a-month plan tops out near $2,000 of API-priced tokens, but they found the plans hand over far more. Since a subscription's margin tracks average utilisation, heavy users drag it well below the roughly 75% gross margin the labs earn selling raw API access.

Xiaomi's MiMo team open-sourced MiMo-Code, an MIT-licensed terminal coding agent forked from OpenCode. Instead of compacting context as sessions grow, it runs a separate checkpoint-writer subagent that logs decisions to a persistent store, so the agent rebuilds state rather than forgetting. In Xiaomi's own double-blind test with 576 developers the two tied under 200 steps, but MiMo-Code won over 65% of longer runs. The benchmarks are self-reported and only against Claude Code.

agentsview is a local-first analytics tool, written in Go, that reads session activity across Claude Code, Codex and more than 20 other coding agents. It pitches itself as a 100x faster replacement for ccusage, letting you audit what your agents actually did and what they cost from one local dashboard rather than a cloud service. The project is trending on GitHub with roughly 1,580 stars.

NVIDIA open-sourced SkillSpector, a security scanner that inspects AI agent skills for vulnerabilities, malicious patterns and other risks. As agents increasingly load third-party skills, the supply-chain problem familiar from package managers moves into the skill layer, and a scanner that flags dangerous behaviour before a skill runs is a sensible guardrail. Written in Python, the project is trending on GitHub with around 2,560 stars.

Output-level testing misses the ways agents fail: a confident answer built on empty tool results, or a correct conclusion that skipped verification. Agent-EvalKit, an Apache 2.0 toolkit from AWS, traces the full execution path instead: which tools an agent called, what they returned, and whether the response reflects that data. It integrates with Claude Code, Kiro and Kilo Code, and you describe evaluation goals in natural language.

FluidAudio is a Swift SDK for fully local, low-latency audio AI on Apple devices, offloading inference to the Apple Neural Engine to cut memory use and avoid the GPU entirely. It bundles speaker diarization, transcription via Parakeet TDT v3, and voice activity detection through open MIT and Apache models, integrated in a few lines. Because it targets background and always-on workloads, none of the audio leaves the device.