New Yorker: 'Sam Exhibits a Pattern of Lying' + Anthropic Hits $30B

GitHub shipped its largest secret scanning expansion in months, adding 37 new detectors across 22 providers and extending push protection to 39 token types by default. The headline addition: secret scanning now works inside AI coding agents through the GitHub MCP Server, catching exposed credentials before commits or pull requests. For teams where agents increasingly generate code and open PRs, this addresses a growing blind spot in the credential leak surface.

The New Yorker obtained seventy pages of Slack messages, HR documents, and secret memos compiled by Ilya Sutskever and sent as disappearing messages to fellow board members before the November 2023 ouster. The material alleges Altman misrepresented facts to executives and board members and deceived them about internal safety protocols. One memo begins with a list headed "Sam exhibits a consistent pattern of..." followed by "Lying."

Broadcom will produce future versions of Google's TPU chips and signed an expanded deal giving Anthropic access to approximately 3.5 gigawatts of computing capacity from Google's AI processors. Mizuho analysts estimate Broadcom will earn $21 billion in AI revenue from Anthropic in 2026 and $42 billion in 2027. Anthropic's run-rate revenue has crossed $30 billion, up from roughly $9 billion at end of 2025.

OpenAI and Anthropic are racing toward potentially record-breaking IPOs by end of year. The WSJ's inside look at both companies' pre-round financials reveals their shared vulnerability: soaring costs to train each new generation of AI models. The timing pairs with Anthropic's revenue crossing $30B and OpenAI's executive reshuffling, making the financial picture more complex than headline growth numbers suggest.

OpenAI's longtime COO Brad Lightcap is moving to special projects, reporting directly to Sam Altman, with his main focus a joint venture with PE firms to sell AI software to enterprises. CRO Denise Dresser takes over some COO duties. CMO Kate Rouch is stepping down for cancer recovery and AGI CEO Fidji Simo is taking medical leave. The reshuffling comes as OpenAI prepares for a potential IPO at an $852 billion valuation.

A detailed GitHub issue with data from months of production logs reports that Claude Code became significantly less reliable for complex engineering tasks after February 2026 updates. The analysis links the regression to reduced "thinking" tokens, showing lower code quality, more errors, and degraded tool use. Read-to-edit ratios dropped and reasoning loops increased, with some users reporting 8-16x higher API costs to achieve the same outcome through retries.

Proxy-Pointer RAG combines the structural reasoning accuracy of Vectorless RAG via PageIndex with the scalability and lower costs of traditional vector retrieval. The hybrid approach integrates PageIndex principles into a vector system for faster retrieval across large, complex document sets. Results show higher accuracy than standard vector RAG without the expensive indexing step that makes full Vectorless RAG impractical at scale.

Meta's data pipeline spans four repositories, three languages, and over 4,100 files. When AI coding agents tried to make edits, they failed because the codebase required tribal knowledge that lived only in engineers' heads. Meta's fix: a swarm of 50+ specialised AI agents that pre-computed 59 context files encoding non-obvious patterns and design choices. Coverage went from 5% to 100% of modules, and preliminary tests show 40% fewer agent tool calls per task.

A Chroma study tested 18 frontier models and found every single one performed worse as input length grew, with some dropping from 95% to 60% accuracy past a threshold. This contradicts the common assumption that more context is always better. The ByteByteGo guide covers strategies for structuring context to avoid these blind spots, including how to manage the tension between comprehensiveness and the degradation that comes with longer inputs.

InfoWorld draws a pointed parallel between the current multi-agent hype and the microservices era, where teams broke working applications into distributed systems then built entire platform teams to manage the complexity they created. Even Anthropic and OpenAI recommend starting with the simplest possible approach. The warning: agent demos with a planner, researcher, coder, and reviewer when a single well-prompted model would suffice are repeating the same mistake.

The US struck over 3,000 targets in the first week of the Iran conflict, made possible by AI-enabled targeting. CENTCOM insists humans remain in the loop, but an Israeli Lavender system operator described spending 20 seconds per target with "zero added-value as a human." The same pattern already exists in business: a Cigna physician denied 60,000 insurance claims in one month, spending 1.2 seconds each. The human is in the loop. Humanity is not.

The median public SaaS valuation compressed 73% from 18.6x revenue in 2021 to 5.1x at end of 2025, while most underlying businesses kept growing. HubSpot grew revenue 141% from $1.3B to $3.1B yet its stock fell 71% from peak. The market is pricing in that classical SaaS, built on the assumption that humans manually enter structured data into fields, faces an existential challenge from AI that can automate the input layer entirely.

Across 4,395 VC financings totalling $186 billion in 2025, vertical AI startups captured 53% of deal volume. Strip out the 12 mega-rounds above $1 billion and vertical companies took 51% of capital too. Healthcare and financial services led with nearly 1,100 deals combined, while manufacturing saw 41% growth from Q1 to Q4. On exits, vertical companies captured 56% of $234.9 billion in total exit value.

RightNow AI released AutoKernel, an open-source framework that applies an autonomous LLM agent loop to GPU kernel optimisation. The agent mimics an expert's workflow: profile, edit, benchmark, keep or revert, repeat overnight. It supports Triton and CUDA C++ backends with a five-stage validation harness. On KernelBench, even frontier models matched PyTorch baseline in fewer than 20% of one-shot attempts. AutoKernel was built to close that gap.

Freestyle provisions full Linux VMs in under 700 milliseconds, clones running VMs without pausing them, and hibernates with exact state restoration. Unlike containers, these are real VMs with root access and full KVM support, meaning agents can run Docker, nested VMs, or any virtualisation stack inside. Built specifically for the scale that coding agents demand, where ephemeral environments need to spin up and tear down thousands of times per day.

Free, open-source macOS app that runs speech-to-text entirely on device using WhisperKit and Qwen 2.5 for cleanup. Hold Control to record, release to transcribe and paste into any text field. No cloud APIs, no data leaves the machine. The local LLM removes filler words and handles self-corrections automatically. Models download once at roughly 3.5 GB total and run on Apple Silicon. The developer notes it offers for free what other apps raised $80M to build.