Older posts.
Model Collapse in 2026: The Loop Nobody Runs, a 40% Synthetic Phi-4, a 49.9% AI Web, and 300T Human Tokens Left
What the model-collapse research tested versus the headline: Shumailov's Nature loop deletes real data every generation and takes OPT-125m from 20 to 28 perplexity, while accumulating data bounds the error; my 100-generation Gaussian experiment shows median variance falling from 0.98 to 0.36 under replacement and holding at 0.98 under accumulation; how Phi-4 (40% synthetic), Llama 3, DeepSeek-R1 (800k samples), Qwen3, Kimi K2 and Nemotron-CC-v2 (38%) actually use it; Epoch's 300T-token stock, Graphite's 49.9% AI web, Cloudflare's 38,065:1 crawl ratio and the $1.5B Bartz settlement.
15 minOct 5, 2026The 2026 SaaS Selloff: IGV Fell 36%, Then Rose 45%. Revenue Grew 11% to 37% the Whole Time
The SaaSpocalypse checked against filings: Anthropic's Cowork plugins on 30 January, Thomson Reuters' worst day ever on 3 February, $285 billion gone in a day and nearly a trillion in a week, a second leg on Mythos in April, IGV down 36 percent to its 10 April low and up 45 percent since; against that, Q2 2026 revenue growth of 11 percent at Salesforce, 24 at ServiceNow, 36 at Datadog and 37 at Snowflake, median retention flat at 110 percent, median multiples from 4.9x to 3.0x to 4.2x, 4,000 Salesforce support roles, Klarna's reversal and $65 billion of Anthropic run-rate.
13 minOct 5, 2026Policy as Code for AI Agents in 2026: Cedar at the Gateway, Rego in the Hook, and the 11.2% the Model Still Lets Through
What actually enforces limits on AI agents in 2026: OPA 1.21 and Rego v1 after the founders moved to Apple, Cedar in Amazon Bedrock AgentCore Policy (GA 3 March 2026, 13 regions, default deny), Kyverno 1.19 retiring ClusterPolicy for CEL, Claude Code hooks that deny even in bypass mode, Codex's Seatbelt and bwrap sandboxes, the MCP 2026-07-28 authorization spec, and why none of it is optional while browser-agent injection success sits at 11.2% and OpenAI says the problem is unlikely to ever be fully solved. With a tested Rego policy and a Cedar policy for tool calls.
13 minOct 5, 2026Kubernetes as the Agent Control Plane in 2026: Agent Sandbox Hit v1.0, 300 Claims a Second, and DRA in Every Supported Release
What shipped for running AI agents on Kubernetes by October 2026: the kubernetes-sigs Agent Sandbox project from a KubeCon preview in November 2025 to v1.0.0 in August, its four CRDs and gVisor, Kata and Firecracker runtime classes, Google's 300 claims per second and 16x growth figures, a density benchmark of 61 Kata agents versus 88, 133 and 274 per node, DRA going GA in 1.34 and locked on through 1.37, kagent, Dapr Agents 1.0, agentgateway, the Inference Extension's move into llm-d, the CNCF survey's 66 percent, and the Gartner 80 percent platform-team prediction nobody has measured.
13 minOct 5, 2026Observability in 2026: OpenTelemetry Graduated, the Collector Is Still v0.162, eBPF Instrumentation Is v0.14, and Datadog Bills $1.12B a Quarter
The non-AI state of observability in 2026 from primary sources: OpenTelemetry graduated from CNCF in May with 12,000 contributors from 2,800 companies, yet the Collector distribution is v0.162.0 with mixed component stability, profiles are alpha, messaging conventions are still Development and OBI, the donated Beyla, is v0.14.0. Datadog grew 36 percent to $1.12B in Q2, cost is the top concern for 31 percent of 1,363 surveyed engineers, and the levers that move the bill (tail sampling, OTTL filtering, cardinality limits, columnar storage, Arrow transport) live in the pipeline you run yourself.
13 minOct 5, 2026MLOps in 2026: Evals Are the New Unit Tests, Judges Agree With Experts 66% of the Time, and OpenAI's Evals API Shuts Down on 30 November
The 2026 state of MLOps, checked against primary sources: OpenAI is closing its Evals API (read-only 31 October, off 30 November) and winding down fine-tuning while buying Promptfoo; Langfuse went to ClickHouse and W&B to CoreWeave; GDPval's model grader agrees with experts 66% of the time against 71% for humans; ARC-AGI-3 went from 0.51% to 99.9% in five months; Terminal-Bench retires a task when every frontier model solves it 5 of 5 times; MLflow 3.16, Kubeflow 26.03, Feast 0.66, lakeFS under BSL, and a pytest gate on a Wilson lower bound that I actually ran.
14 minOct 5, 2026MCP in 2026: 475M SDK Downloads a Month, 39,492 Servers in a Registry Still in Preview, and 12 Gateways Selling the Same 3 Features
The MCP ecosystem ten months after the Linux Foundation took it over, measured from primary sources: 475 million SDK downloads a month across npm and PyPI, three spec revisions ending in the stateless 2026-07-28 rewrite, an official registry I paged to 39,492 servers (24,242 remote) that is still labelled preview while Glama lists 96,340, every major client with different controls, twelve gateways from $0 to $0.005 per thousand calls, the tool-overload numbers (55k tokens before the first prompt, 85 percent recoverable), and a stateless Python server on mcp 2.3.0 that I ran and tested.
14 minOct 5, 2026Infrastructure as Code in 2026: Two Forks at 1.16 and 1.13, a $6.4B Owner, 912 Public State Files, and One Agent That Ran terraform destroy
Infrastructure as code three years after the BSL relicence: Terraform 1.16.5 under IBM versus OpenTofu 1.13.1 under the Linux Foundation and who shipped what first, HCP Terraform at $0.10 to $0.99 per resource with the legacy free plan gone, CDKTF and System Initiative archived, Pulumi 3.267 and Crossplane 2.4, a Terraform MCP server that grew from registry lookups to workspace administration in 15 months, 44 percent running AI for infrastructure but 34 percent trusting it, 912 exposed state files with 41 live AWS keys, and the agent that ran terraform destroy on 2.5 years of production.
13 minOct 5, 2026GPU Clouds in 2026: H100s at $2.59 an Hour, $104B of CoreWeave Backlog on $35B of Debt, and a Six-Year Bet on Silicon
The GPU rental economy from filings and price pages fetched on 5 October 2026: H100 hours that fell from about $8 in 2023 to a $1.70 trough in October 2025 and recovered to $2.59 on the transacted index while list prices rose 25 percent; CoreWeave's $2.58B quarter, $103.7B backlog, $35.1B of debt from 15 percent DDTLs to a 5.9 percent A3 facility, and Microsoft at 67 percent of 2025 revenue; Nebius, Lambda, Crusoe, Oracle's $664B RPO and NVIDIA's $105B of guarantees; who books servers over 5, 5.5 or 6 years and what Burry's $176B means; and an own-versus-rent break-even calculator.
13 minOct 5, 2026FinOps for AI in 2026: 98% Manage Token Spend, 3 in 4 Can't Prove the Value, GPUs Run at 5%, and FOCUS Gets a Token Column in 1.5
The AI cost ledger as of October 2026: State of FinOps 2026 (1,192 practitioners, 98 percent manage AI spend, up from 31 percent in 2024), the Tokenomics Foundation finding that three in four enterprises cannot prove AI outcomes to the CFO, a per-million-token price table for Anthropic, OpenAI, Google and DeepSeek with cache and batch multipliers, a 20-step agent loop at $1.58 uncached and $0.42 cached, the FOCUS 1.2 to 1.5 roadmap and who exports which version, LiteLLM and Cloudflare budgets, Cast AI's 5 percent GPU utilisation, and $0.0037 of model cost inside a $0.99 resolution.
13 minOct 5, 2026CRA Reporting Went Live on 11 September: 24 Hours to ENISA, 3.08 Billion Rekor Entries, 17% of PyPI Attested, 92% SBOM False Positives
Supply-chain compliance in 2026: what the Cyber Resilience Act's Article 14 duty requires since 11 September 2026 (24-hour early warning, 72-hour notification, ENISA's Single Reporting Platform), what waits until 11 December 2027, how the open-source steward role works, where SBOMs stand (CISA's 2026 minimum elements, CycloneDX 1.7, a 92 percent false-positive rate in a 2,414-repo study), how far provenance has got (SLSA 1.2, 20 percent of PyPI uploads via trusted publishing, Rekor at 3.08 billion entries), the US retreat from mandates, and the pipeline I would run.
13 minOct 5, 2026DevOps Still Matters in 2026: AI Cut Delivery Stability 7.2%, Then Doubled Merged PRs and Added 91% to Review Time
Why DevOps is the big thing of the agent era: DORA 2024 found a 25% rise in AI adoption cost 1.5% throughput and 7.2% stability, DORA 2025 saw throughput turn positive while instability stayed up across nearly 5,000 respondents, and DORA's 2026 ROI model budgets a 15% three-month dip and a change failure rate rising from 5% to 6%; GitHub merged 518.7M PRs (+29%) and over 1M agent PRs in five months, Faros telemetry on 10,000 developers shows 98% more PRs and 91% longer reviews, METR found experienced developers 19% slower, and the Replit postmortem's fixes are 2015 DevOps controls.
13 minOct 5, 2026Coding Agents in 2026: Cursor at $2B Then Sold to SpaceX, Cognition at $1B, 58% on Terminal-Bench, and the Only RCT Still Says Slower
The coding-agent market as the primary sources report it in October 2026: Cursor from $500M to $1B to $2B in nine months and then acquired by SpaceX, Cognition from $73M to a $1B run-rate at a $48B valuation, Claude Code past $500M, Lovable at $13.3B, nearly 140,000 organisations on Copilot; four re-pricings from requests to tokens; Terminal-Bench 4.0 topping out at 58.2 percent for $3,267 a run; METR's redesigned study at minus 4 to minus 18 percent against 1.4 to 2x self-reports; 45 percent insecure samples, 1.7x more issues per AI pull request, and curl's bug bounty closed.
14 minOct 5, 2026CI/CD in 2026: Agents Opened 1M PRs in 5 Months, Bots Write 1 in 5 Reviews, and Most Agent PRs Get No Human Look
What the pipeline looks like when AI agents open the pull requests: GitHub's coding agent went GA on 25 September 2025 and opened over a million PRs in five months, Copilot code review passed 60 million reviews and one in five on GitHub, CodeRabbit has reviewed 13 million PRs on $88 million raised, and the first studies find most agent PRs get no human review attention. The gates that still hold: the assigner cannot approve, an extra approval for bot authors, path fences, signed commits, SLSA provenance, and a workflow YAML that gives agent PRs their own lane.
13 minOct 5, 2026Browser Agents in 2026: 90.2% on OSWorld-Verified, 20.6% on OSWorld 2.0, and Three Agent Browsers Already Shut Down
Computer-use agents passed the 72.36 percent human baseline on OSWorld-Verified this year (90.2 percent for the top framework, 86.0 for Claude Fable 5), then scored 20.6 percent on the 108-task OSWorld 2.0 that replaced it; Operator, Project Mariner and ChatGPT Atlas were all discontinued within 19 months; prompt-injection success in Anthropic's own tests fell from 23.6 to 11.2 percent and no vendor claims zero; Cloudflare sells HTTP 402 access to agents while 52 percent of crawling is for training; and hosting settled on one headless Chrome per microVM at $0.02 to $0.12 a browser-hour.
13 min