Our point of view on shipping AI — page 3 of 5.
- AI Engineering6 min
LLM Testing Already Had a Playbook in 2020
By Petru Popa ·Teams treat testing a language model as a new discipline with no precedent. The methodology that fits it best was published in 2020, and the test type it names as most valuable is the one almost nobody writes.
Read - AI Engineering6 min
Fine Tuning vs RAG Is the Wrong Question
By Petru Popa ·Two papers settle the versus framing between them. One finds retrieval beats unsupervised fine-tuning at teaching a model facts. The other finds the gains from both are cumulative — and things that add are not substitutes.
Read - AI Strategy6 min
The Watermark Is Sparsest Where You Would Audit
By Petru Popa ·Anthropic will watermark future Claude models using a version of SynthID-Text. The announcement and the Nature paper behind it agree that the mark weakens on code, short answers, and factual passages, which is where most enterprise detection requirements are aimed.
Read - AI Engineering6 min
Verified Is Not Reproduced
By Petru Popa ·A community effort reproduced claims from 2,226 ICML 2026 papers and reported 51% with at least one verified claim. Opening the frozen verdicts dataset shows what that percentage covers, what a $20 compute cap excludes, and why there is no way to look your paper up.
Read - Dev Tools6 min
Conformance Lets a Client Ignore Half Your Plugin
By Petru Popa ·GitHub shipped Agent Plugins 1.0 across VS Code, Copilot CLI, the Copilot SDK and the Copilot app. The spec repository behind it standardizes the package and explicitly defers trust, permissions, sandboxing and provenance to the client — and lets a conforming client implement skills without MCP servers.
Read - Dev Tools6 min
The Audience Is the One Claim Your Attacker Picks
By Petru Popa ·A proposal to let GitHub Actions workflows pre-declare OIDC audiences is the wrong fix to wait for. Reading the permission schema and AWS's condition-key reference shows why the control that holds today lives on the relying party.
Read - AI News6 min
The 24 GB Card Fits the Weights, Not the Features
By Petru Popa ·Meta's Muse Glimmer targets consumer GPUs with 24 to 32 GB. Reading config.json and the GGUF file listing, the binding constraint on a 24 GB card is not the 131,072-token context — it is the vision projector and the speculative-decoding drafter, the two files that are the reason to pick this model.
Read - AI Engineering5 min
The 3.3x Belongs to the Model, Not the Flag
By Petru Popa ·vLLM published a 3.3x throughput gain for decode context parallelism on Kimi K2.6. The config source shows the ceiling for grouped-query models is set by KV head count — and for Qwen3-235B-A22B on one node, that ceiling is two.
Read - Dev Tools6 min
Three Meters and One Brake
By Petru Popa ·Between 6 and 7 August, GitHub shipped per-agent usage metrics, an effort dial for Copilot code review, an ROI section that prices a developer in pull requests per month, and an organization-wide pull request limit. The limit is documented as applying only to public repositories and only to users without write access — which is not where your agents are.
Read - Dev Tools6 min
The Client List Is the Policy
By Petru Popa ·MCP allowlists went generally available in enterprise managed settings, and a day later enterprises could install third-party GitHub Apps. GitHub's own reference pages name the clients that enforce the first and the missing event stream under the second, and the Copilot cloud agent is covered by neither. What that changes about how you govern agents this quarter.
Read - Dev Tools5 min
MCP's Twelve-Month Deprecation Window Doesn't Cover This Revision
By Petru Popa ·MCP revision 2026-07-28 makes the protocol stateless and adopts a feature lifecycle policy with a twelve-month floor before removal. The removals that will cost you time this quarter never entered that policy, and the deprecated feature with the nearest removal date is the one that has read as deprecated since March 2025.
Read - AI Strategy6 min
The 50% Circuit Breaker Counts Disclosures, Not Code
By Petru Popa ·Rust now has a written policy for LLM-assisted contributions to rust-lang/rust, including a threshold that suspends merges when more than half of PRs in a six-week window are LLM-created. Read what the threshold counts and the policy stops looking like a quality control.
Read