Development

  • InfoQ — Cloudflare launched K2 in public beta, a serverless event-streaming service built on R2 object storage instead of a Kafka-style replicated log, claiming ~1s p99 produce latency; a beta tester measured 2.5s p95 / 7.5s p99 end-to-end, while competitor s2.dev claims ~50ms p99 with a stateful-flush design.

AI

  • Simon Willison’s Weblog — Simon Willison released ttok 1.0, switching its default tokenizer from GPT-4’s to the newer GPT family’s, based on third-party evidence (unconfirmed by OpenAI) that GPT-6 shares that tokenizer.
  • arXiv cs.CL — A meta-analysis of 3,653 models across 97 datasets finds automated verbal deception detection plateaus at 74.4% accuracy (95% CI 71.2–77.4%) regardless of model sophistication; newer LLM/embedding approaches don’t beat it, and only 12.46% of studies used verifiable ground truth.
  • arXiv cs.CL — Auditing six frontier LLM judges across four benchmarks finds verdicts are non-reproducible even at temperature zero, and swapping presentation order flips most verdicts on hard tasks; the most “consistent” judge was just consistently wrong, matching ground truth only 51% of the time — rubric scoring beat pairwise comparison at fixing it.
  • arXiv cs.CL — “Thinking Inertia”: explicit “don’t show reasoning” instructions don’t reliably suppress visible chain-of-thought in LLMs, especially as the answer space opens up past boolean/multiple-choice; forcing answer-only compliance trades off against accuracy on open-ended tasks.
  • arXiv cs.CL — Testing AI-text detectors on 4,000 pre-ChatGPT PNAS abstracts rewritten by 23 LLM versions (3 vendors, June 2023–Aug 2026) finds detection accuracy collapses at model-version boundaries: one detector caught 99%+ of rewrites just before a boundary but only 3.8% just after — benchmark accuracy is “provisional” and breaks with every new model release.
  • arXiv cs.CL — “TRACE” shows agent-benchmark score changes can come from the evaluation harness, not the agent: renaming tools alone dropped a scripted, capability-unchanged agent’s score by 0.250, and identical reruns flipped 15–36% of task outcomes — enough noise to swamp real capability differences.
  • arXiv cs.CL — “SWE-Journey” finds coding assistants pass over 75% of requested-functionality tests when the user is a software architect but under 25% when the user is a non-coder, arguing current assistants don’t yet reliably serve non-coders.