https://keep-up.ai A weekly AI & LLM news magazine — curated papers, tools, and industry updates for ML engineers. en-us https://keep-up.ai/digest/2026-w30 https://keep-up.ai/digest/2026-w30 Tue, 21 Jul 2026 08:53:31 GMT The cheapest reading of the week is "open weights caught up." Kimi K3 lands near Opus 4.8 on Artificial Analysis's Intelligence Index and takes #1 on Frontend Code Arena at a fraction of proprietary pricing. The more useful reading is what "caught up" now means for labs that don't own their own power. Wojciech Gryc's economics piece makes the tension explicit: when variable costs scale with revenue and the top slot trades every six weeks, the only durable moats are infrastructure, product, or a regulatory ceiling — and Anthropic is the lab most exposed on all three. Against that, the week's safety disclosures land differently. GPT-Red reports 84% attack success against GPT-5.1 on an indirect prompt injection benchmark, then folds those attacks back into GPT-5.6 training — a self-play flywheel gated by compute that only the largest labs can spin. Meanwhile Ayush Paul walked Claude's web_fetch, letter by letter, through a honeypot's embedded links to exfiltrate user data — a primitive-shaped hole in a lab already running that flywheel. Robustness gains inside the training loop are real; the gap between them and what ships is not closing at the same rate. Jeremy Theocharis's Local-First Conf essay is a third data point worth reading against both. He documents ~$10k/month in token spend, agrees with most critics on the substance, and uses the tools daily anyway. That dissonance — heavy use paired with honest acknowledgment of the failure modes — is closer to the median staff-engineer experience than most adoption narratives, and it's what makes the two safety stories above matter in practice rather than in principle. https://keep-up.ai/digest/2026-w28 https://keep-up.ai/digest/2026-w28 Mon, 13 Jul 2026 00:00:00 GMT Cost-aware tiering showed up at every layer of the stack this week — API family, runtime, orchestration wrapper, log analyzer, reward model — and it's tempting to read that as consensus finally arriving. The more interesting read is that the layers disagree on what "cheaper is fine" actually means. The frontier-plans-cheap-executes pattern lands as a clean API primitive at one end of the week and as a calibration problem at the other. OpenAI ships it as a menu (Sol/Terra/Luna, plus a four-agent parallel mode). Modal argues the runtime beneath that menu has to be rebuilt from Kubernetes-era assumptions. The community tools show the same pattern reached from the opposite direction — Pilotfish routes by role, Frugon audits by log — and both hedge exactly where the vendor pitch doesn't: Pilotfish deliberately keeps security work off the frontier model because safety classifiers refuse legitimate defensive work, and Frugon flags in its own output that "within tolerance" is an offline estimate you should validate against real traffic. The citation-verifier paper is the one landing the sharpest version of the same point. Its finding isn't "smaller judges work" — it's that judges with statistically indistinguishable F1 differ substantially in false-positive and false-negative rates, and that directional bias is what a reward loop reinforces. Read alongside OpenAI's own mid-week SWE-Bench Pro post-mortem, which questions the scoreboard where Claude leads it, the cheap tier isn't the risky decision this week. The scalar metric people use to justify either tier is. https://keep-up.ai/digest/2026-w25 https://keep-up.ai/digest/2026-w25 Wed, 17 Jun 2026 00:00:00 GMT The operational layer around agents got a lot of attention this week — meta-harnesses, persistent cloud execution, engineering-org rules calibrated for autonomous execution. Read together, they describe a stack that is finally being built out above the model. Read against Anthropic's Friday-night Fable/Mythos suspension, they describe a stack whose floor can drop away on a phone call. Will Larson's post is the most concrete artifact of what running at the upper layer actually looks like: an entire year of migrations, deploy pipelines, and process redesign predicated on the assumption that the base case of most work is now agentic. Databricks and OpenAI both shipped things this week that would make that easier for teams operating one tier lower — a common interface across Claude Code, Codex, Cursor, and Pi; a persistent cloud execution substrate for Codex sessions that outlive a laptop lid. The bet is that harnesses will multiply and models will churn, but the layer above them stays. The Fable/Mythos suspension is the argument against that bet in its purest form. Artificial Analysis's Intelligence Frontier chart moved backward for the first time. Nothing in Omnigent's YAML or Codex's persistent sandbox reroutes around an export-control directive that revokes access overnight for a vendor sitting on top of the coding leaderboard. The meta-harness answer to model risk — swap providers with a one-line change — is a real answer, but it presupposes you have a competitor at parity to swap to. This week that assumption got weaker, not stronger. https://keep-up.ai/digest/2026-w16 https://keep-up.ai/digest/2026-w16 Sun, 19 Apr 2026 00:00:00 GMT Two Anthropic stories this week are really one story told from opposite ends. Opus 4.7 shipped with real-time cyber safeguards explicitly described as a testbed for what Anthropic hopes to eventually do with Mythos-class models — the model they restricted last week after it autonomously chained zero-days across major OSes and browsers. Read together, the launch post reads less like a capability announcement than an operational admission: we can now ship a model only because we have learned to differentially suppress parts of it at training time and intercept prohibited use at inference time. "Same pricing as 4.6, state-of-the-art on CursorBench" is the surface; the sub-narrative is that the frontier now routinely produces capabilities that need to be partially unlearned before release. The tension sharpens when you stack the launch against its own footnotes. Anthropic concedes in the migration guide that the new tokenizer consumes 1.0–1.35× more tokens for equivalent input, and that higher effort tiers think longer on later turns. The community benchmark claiming a ~50% effective cost increase and a sharp MRCR 256K regression is not a contradiction of the launch post — it is a literal reading of it. The coding gains are real. So is the bill. Against that backdrop, the week's quieter items — OpenAI's SDK picking up native sandboxes and MCP, Raschka arguing that working reference code beats technical reports for understanding architectures — both point the same direction. Model internals are getting less legible from papers, while the scaffolding around models is getting more standardized. Where you can still see what is happening, look there.