cache it, weigh it, x-ray it, and keep it local
Week ending February 20, 2026; archive coverage: February 19.
Some weeks the news is a fireworks show. This week it was a maintenance crew. No consumer demo I care about, no funding round worth your attention, just a handful of people quietly fixing the parts of the stack that actually determine whether your agent is fast, cheap, honest, and yours.
That's the good kind of week. The layer under the models is where the real leverage lives, and four different groups spent it there: one on latency and cost, one on open agentic coding, one on security that doesn't take the model's word for it, and one on running speech-to-text without renting it.
Let me take them in order.
caching is a systems problem wearing a prompt costume
The OpenAI cookbook on Prompt Caching 201 is the most useful thing I read all week, and it's basically a plea to stop sabotaging yourself. The mechanism is simple: the model reuses the key/value tensors it already computed for a repeated prefix, so it skips prefill. Up to 80% off time-to-first-token, up to 90% off input tokens, no extra fee.
The catch is that caching wants stability and everything else you do wants change. Cache hits require an exact prefix match, in 128-token increments, over 1024 tokens minimum. A timestamp near the top, a reordered tool schema, a nudge to reasoning effort, and you've quietly nuked your own hit rate. Their advice is unglamorous and correct: put durable stuff (instructions, tools, schemas) first, shove volatile stuff last, move debug metadata out of the prefix, and use prompt_cache_key to keep related traffic sticky to the same machine. One coding customer went from 60% to 87% hits just by setting that key.
The part worth internalizing: context engineering and caching are at odds. Compaction and summarization save tokens but bust the cache. So decide that trade with evals, not vibes.
an open model that says it grew up
The GLM-5 technical report frames itself as moving from "vibe coding" to "agentic engineering," which is exactly the kind of slogan I'd normally roll my eyes at. But the engineering claims are substantive. They lean on DSA to cut long-context training and inference cost, and they built an asynchronous RL setup that decouples generation from training, plus async agent RL algorithms aimed at long-horizon interactions.
That decoupling matters more than any leaderboard number. Long-horizon agent RL is bottlenecked by rollouts, and if generation stalls your trainer, you burn GPUs waiting. State-of-the-art-on-open-benchmarks is a claim to verify, not to swallow. But code and weights are on GitHub, so at least it's checkable, and an open model targeting end-to-end software work is a good thing to have around.
stop asking the model if it's being naughty
Zenity's maliciousness classifier built on LLM internals is my favorite kind of skeptical. Instead of trusting input/output surfaces or asking a model to grade itself, they feed the prompt through Llama-3.1-8B, grab the activations, and run a cheap logistic-regression probe. Prompting the same model to self-classify does worse across every category, which they attribute to the model understanding something is off but failing to articulate it. Read the internals, not the confession.
The real gift is the evaluation discipline. They hold out an entire dataset at a time, leave-one-dataset-out across 18 open datasets, so the test data is genuinely unseen. Standard train/val/test accuracy looks great; the honest out-of-distribution number is much harshÂer. "Don't be an optimist when it comes to security" is a line I'm stealing. The false-positive rate is still too high to run standalone, and they say so, which is why I trust the rest.
dictation you can audit
Finally, Amical, an MIT-licensed, local-first dictation app running Whisper and open LLMs on-device. It's context-aware, formatting for Gmail versus Slack versus your IDE versus a terminal. The pitch is simple: your voice never leaves the machine by default, one-click model setup instead of hand-wiring Whisper, and MCP integration on the roadmap. Against $15/month subscription tools, owning the thing outright is the whole point.
what I'd do Monday
Instrument cached_tokens before you touch anything, then stabilize your prefix and set a prompt_cache_key scoped to stay under ~15 RPM per key. Re-run one of your safety evals leave-one-dataset-out and see how much of your accuracy was self-deception. And if you handle anything sensitive, try a local-first dictation setup this week instead of piping your voice to a vendor by default.