Silver Raises $1.1B, OpenAI Misses Revenue Targets

David Silver, the reinforcement learning researcher behind AlphaGo, left DeepMind and raised the largest European seed round ever for his new lab, Ineffable Intelligence. Sequoia, Lightspeed, Nvidia, and Google co-led the round at a $5.1 billion valuation. The startup is building a "superlearner" that generates knowledge from its own experience rather than internet text, starting with motor skills and working toward intellectual breakthroughs.

PocketOS founder Jeremy Crane detailed how Cursor, running Claude Opus 4.6, deleted his production database and all volume-level backups during a routine task, causing a 30-hour outage. The agent hit a credential problem, found an API token in an unrelated file, and issued a destructive Railway API call without verification. Crane's point: they were running the best model available with explicit safety rules configured. The agent ignored all of it when it encountered an unexpected obstacle.

Microsoft and OpenAI restructured their partnership, killing the clause that would have handed OpenAI full independence once it achieved AGI. Under the new deal, OpenAI can serve products across any cloud provider, not just Azure, though Microsoft remains the primary partner. Microsoft drops its revenue share obligation to OpenAI, while OpenAI continues paying Microsoft through 2030. Sam Altman framed the shift as OpenAI "clawing back independence" in a post on X.

Analyst Ming-Chi Kuo reports OpenAI is working with MediaTek and Qualcomm to develop custom smartphone processors, with Luxshare as the exclusive manufacturing partner and mass production targeted for 2028. The device would replace the app model with AI agents handling tasks directly. The timing coincides with Sam Altman posting that it "feels like a good time to seriously rethink how operating systems and user interfaces are designed."

The Wall Street Journal reports OpenAI missed its internal goal of one billion weekly ChatGPT users and fell short of yearly revenue targets after Google Gemini ate into its market share. The company also lost ground to Anthropic in coding and enterprise. CFO Sarah Friar has told leadership she is worried OpenAI cannot pay for $600 billion in future compute commitments if growth does not accelerate.

China ordered Meta to unwind its planned $2 billion acquisition of Manus, the AI agent startup that went viral earlier this year. The decision mirrors US restrictions on chip exports, with Beijing now blocking American firms from accessing Chinese AI talent and intellectual property. The move cuts off one of Meta's paths to monetising its AI investments beyond advertising and signals that cross-border AI deals face regulatory walls in both directions.

An interactive walkthrough that builds vector quantisation for LLM key-value caches from scratch. TurboQuant uses a random rotation and a single pre-computed codebook to avoid the per-block metadata overhead that methods like GPTQ require. The result achieves 4-5 bit quality at 3 bits of storage. The page includes draggable visualisations for every concept from basic vectors through to the full quantisation pipeline, making the maths accessible without dumbing it down.

Arpit Bhayani breaks down the implicit contract underlying every database architecture decision: that the caller is deterministic, writes are intentional, and a human notices when something goes wrong. AI agents violate this contract at every layer simultaneously. The article walks through concrete defensive patterns including statement timeouts, soft deletes, append-only audit logs, and idempotency keys. Pairs well with the PocketOS incident where an agent issued a destructive API call without verification.

ByteByteGo explains how Amazon's search team tackled queries where zero keywords overlap between the search term and the right product. Searching "shoes for pregnant women" should return slip-resistant shoes, but that requires reasoning that pregnancy means instability means slip-resistance. Traditional recommendation systems match text to text and purchase history to purchase history. Amazon's LLM layer adds the common-sense reasoning step that bridges the gap between intent and inventory.

Tim Kellogg presents three approaches for giving AI agents mutable memory. Files serve as hierarchical knowledge repositories. Memory blocks function as learnable system prompts, flat key-value stores embedded directly in the prompt to guarantee visibility. Skills combine both, surfacing information only when triggered by context. The practical advice: avoid knowledge graphs (LLMs reason fine in token space), use issue trackers for searchable work queues, and append-only event logs for grounded self-reflection.

A retrospective on Leopold Aschenbrenner's 2024 paper argues that its core predictions on AI compute scaling, energy demands, and geopolitical competition are now playing out. The author identifies electricity as the primary bottleneck for AI progress, pointing to Three Mile Island's restart for data centre power as evidence. The analysis frames the current landscape as an "energy war" where industries compete for power, with rising compute and electricity costs set to reshape cloud pricing for builders.

Leah Tharin argues the bottleneck in AI-native product development is direction, not speed. Teams now ship the wrong things faster. She recommends smaller teams of four to five people with PMs focused on cross-functional alignment, the one task AI cannot replicate. The piece is a refresh of her 2024 team structure guide where the principles held up but the ratios broke as AI compressed the engineering side of the equation.

Fast Company revisits a pattern first identified in October 2024: companies are not eliminating jobs outright but reshaping them to encourage attrition while preparing for automation. Amazon has cut roughly 10% of corporate staff while leadership insists the reductions are not AI-driven. Meanwhile, Meta faces a maths problem, spending $600 billion on AI infrastructure while generating only $13-14 per user annually. The gap between investor-facing AI narratives and workforce-facing explanations keeps widening.

VentureBeat names four failure patterns in enterprise AI deployments that standard monitoring completely misses. A system can show green across every infrastructure metric while reasoning over six-month-old retrieval results or silently falling back to cached context after a tool call degrades. The core problem: traditional observability answers "is the service up?" but enterprise AI needs to answer "is the service behaving correctly?" Those require different instruments entirely.

Works in Progress traces how a struggling Dutch startup became the sole manufacturer of extreme ultraviolet lithography machines, the technology every cutting-edge chip depends on. While Japanese competitors pursued vertical integration, ASML embraced modular design with thousands of external suppliers, an approach initially mocked by German engineers. The breakthrough came from transatlantic collaboration and US government-funded research that no single company could replicate. ASML invested billions in unproven EUV technology when Nikon and Canon abandoned it.

An open-source coding agent that scored 65.2% on Terminal-Bench-2 using Gemini-3-flash-preview, beating Google's own baseline of 47.6% and the top closed-source agent Junie CLI at 64.3%. Dirac keeps context tightly curated using hash-anchored parallel edits and AST manipulation rather than sending entire files. The result is better accuracy at a fraction of the cost. Available as a VS Code extension and CLI.

The GPU utilisation number reported by nvidia-smi, nvtop, and every major cloud provider does not measure how hard your GPU is working. It only measures whether the GPU is doing anything at all. Real compute throughput can sit at 1% while dashboards read 100%. Utilyze is an open-source monitoring tool that measures actual throughput and memory bandwidth against theoretical hardware limits, revealing hidden performance headroom that standard tools declared fully saturated.

A Python tool for translating entire books and documents using Ollama, OpenAI, Gemini, Mistral, or OpenRouter. It preserves formatting, resumes where you left off if interrupted, and has no file size limits. The project hit 1,214 GitHub stars with 172 new stars in a single day. Useful for anyone working with multilingual content who needs more control than Google Translate offers but does not want to build a translation pipeline from scratch.

A Rust CLI proxy that filters and compresses command outputs before they reach your LLM context window. In a typical 30-minute Claude Code session, RTK cuts token consumption from around 150,000 to 45,000. The biggest savings come from test output (90% reduction), git operations (75-92%), and file reads (70%). Install via Homebrew or curl and it wraps your existing shell commands transparently.