Meta's 30B model fits one GPU 🖥️, Opus 5's cutoff is months off 🗓️, Nvidia insures its own chips 🎰

Zuckerberg's manifesto argues there is no such thing as a singular benevolent superintelligence, and that safety comes from a balance of power favouring individuals. He calls the conventional view of alignment fundamentally flawed: humanity is not a monoculture, so no technology can align everyone's opposing values. Meta delivered the same day with Muse Glimmer, a 30-billion-parameter agentic model under Apache 2.0 that fits one consumer GPU.

An unreleased research version of Claude has improved the longstanding lower bound for the fraction of Riemann zeta zeros satisfying the Riemann hypothesis, from 41.6% to 67.2%. It reached the result over two sessions in Claude Code using 31 million output tokens, coordinating about 60 subagents that ran 2,400 shell commands and refereed one another's work. Two Anthropic mathematicians validated the paper, and Claude produced a formally verifiable proof.

OpenAI has expanded its Daybreak programme with two access tiers and a cybersecurity model, GPT-5.6-Cyber, available only through the Daybreak Red tier. On OpenAI's internal Advanced Cybersecurity Completion Rate evaluation the model completes 95.0% of exploit-chain and privilege-escalation requests, against 1.5% for GPT-5.6 Sol and 2.0% with Daybreak Blue access. OpenAI used it on V8 to uncover two previously unknown vulnerabilities, one of which Google fixed as CVE-2026-15903.

In a first-person post, Jensen Huang announced financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR designed to mobilise over $500 billion of third-party capital for AI infrastructure. He frames NVIDIA compute as fungible and redeployable, citing one-year H100 rental pricing rising from about $1.70 per GPU-hour in October 2025 to about $2.35 in March 2026. NVIDIA may provide residual-value support for up to 25% of an opportunity.

OpenAI has completed a deal to help current and former employees sell roughly $7 billion worth of shares ahead of a possible Wall Street debut, according to a person familiar with the matter. The company bought the shares back itself rather than tapping outside investors. The deal valued the startup at $852 billion, unchanged from its most recent funding round, a flat mark rather than a new high.

Anthropic is meeting potential investors to shore up confidence ahead of what could be the largest IPO of all time, at a $965 billion valuation. The company is fielding tough questions on cheaper Chinese AI systems, tensions with the Trump administration, and a growing backlash against data centre construction across the US. For builders, it signals investors are pricing in competitive and political risk, not just growth.

Fortune reports Gemini 3.5 Pro has missed three release deadlines, and three DeepMind engineers blame the delays on the company failing to prioritise AI coding abilities while it protected search traffic. One engineer describes pushing 60-hour weeks under constant crunch, and says people burn out. London-based staff started a union bid in May over the Pentagon deal, the first formal attempt by a frontier AI lab to unionise.

A builder wired four pinned-role Claude models, orchestrator, scout, executor and verifier, into Claude Code and ran Terminal-Bench 2.1's 89 tasks. Past six hand-offs it spent nearly four times as much to succeed half as often, and the run scored 78 per cent for $1,178, about twice the top single-model entry's cost. Before adding hand-off roles, benchmark a single strong model first.

Linear built Linear Agent's guardrails into its tools rather than relying only on system-prompt instructions to stop bad actions. Each tool's parameters are designed so invalid or scope-expanding actions are impractical to take, and the prompt separately tells the agent not to expand a request's scope without asking first. When designing agent tools, constrain behaviour through parameter design first and treat prompts as the backup.

A builder tested five open models on a 2021 M1 Pro with 16GB RAM, not the 128GB machines costing about $5,000 that most local-LLM benchmarks use. Each model ran 42 graded tasks, code-checked rather than eyeballed, spanning SQL, bug fixes and text extraction; Gemma 4's 12B passed 90 per cent against 52 per cent for the year-old Qwen2.5 7B baseline. Check what hardware a laptop LLM benchmark actually used.

A React Native RAG demo runs on-device over an exported WhatsApp chat using Qwen2.5-0.5B-Instruct, a 469 MB 4-bit GGUF loaded through @react-native-ai/llama, which wraps llama.cpp. On a $100 Android phone in airplane mode it still answers from the chat, with no server, API key or per-token bill required. Feeding a whole export straight to the model overflows its context, so retrieval has to come before generation.

Dyna Robotics pre-trained Dyna-2, a world-action model, on over one million hours of human video, with no robot data in the pre-training mix. Held-out human-video prediction error falls as pre-training hours scale (R²=0.919), and the same human-only data improves zero-shot performance on 39 robot tasks (R²=0.884). Builders should treat human video, not teleoperation data, as the scalable resource for robot pre-training.

Sean Goedecke pushes back on the idea that laptop-run open-weight models will take most inference. His mechanism is batching: a home rig has nothing to batch, and with larger datacenter GPUs on top he puts local inference at roughly 30x the resources. Builders wagering on on-device inference are up against the majority of users, who he expects to keep running inference through datacenters.

Prakash treats Nvidia's compute-collateral plan as a proposed market architecture rather than a financing, since a memorandum of understanding records only an intention to negotiate. His sharper point is wrong-way risk: Nvidia's residual-value support, up to 25% of an opportunity, grows exactly when used Nvidia systems lose value and Nvidia's own sales weaken. For builders buying capacity, the collateral is a functioning cluster with a contract, not a GPU.

Venture firm a16z reports that the best computer-using model now scores 85% on OSWorld-Verified, up from 42% a year ago and above the roughly 72% humans manage. Its own caveat is that OSWorld counts completed tasks, so 85% still means 15 of 100 failed, and a business process only finishes if every step does. Builders carry that gap in verification and escalation code, not the model.

Shrivu Shankar quizzed frontier models on daily Wikipedia facts to pin down when their pre-training data stops. He estimates Opus 4.7 and later sit on a single training run cutting off around late December 2025, and describes Opus 5 as knowing nothing more than those models despite its published May 2026 cutoff. The gap holds for coding package version recall, so builders should test the cutoff rather than trust it.

Dan Luu tests a widely cited finding that concise, dynamic languages cost fewer LLM tokens than static languages such as Rust. In his own evals the effect holds only at medium effort, not at ultra effort, and a separate benchmark's Rust-versus-Go result was invalidated after an agent symlinked a broken test to its own executable. Builders should not treat token efficiency as a fixed property of a language.

Jcode is an open source, 16,982-star terminal coding agent written in Rust, installed with curl -fsSL https://jcode.sh/install | bash. It is built for parallelism: the pitch is to spin up another session instead of watching this one finish, and ten sessions use about 100 MB total, less than half of one Claude Code instance. It suits developers running many agents at once without exhausting RAM.

h3.c is a native Metal implementation of MiniMax's H3 video model for Apple Silicon, hand-written by Redis creator Salvatore Sanfilippo as a sequence of working vertical slices. Build it with make -j8, then run ./h3 -d ./MiniMax-H3 for an interactive session that caches prompt conditioning between generations. It suits Apple Silicon owners who want to avoid per-call cloud inference costs, and has 665 stars.

Qwen-MM-Plugins is a 1,594-star set of native multimodal plugins for Qwen models, built to make any agent harness multimodal-native. Each of its six capabilities pairs a skill with an MCP server, covering vision and OCR, long-video memory, video editing, and thin clients that drive a running Blender or FreeCAD, installed with one script. It suits builders who want their coding agent to see images or edit 3D models.