wireless

stealth | previously ML @ google

[wireless @wireless_dev]

@MParakhin i think claude is especially good at grokking the underlying intent behind the user's instructions, which imo is different from instruction following, and is maybe more related to pretraining, though i dunno

[wireless @wireless_dev]

@_zafaris @samhogan Conductor Cloud

[wireless @wireless_dev]

@_kanarazu_ Probably bigger than people think but will take a while for it to come to fruition, at which point the interesting part of the opportunity (relative edge at doing ZK stuff or whatever) will probably be democratized by LLMs.

[wireless @wireless_dev]

@zeeg Maybe they can't say this out loud but I think labs want to own compaction because their version uses internal reasoning withheld from the API (at least OpenAI's does), thus can't be replicated externally.

[wireless @wireless_dev]

@gakonst @vex_0x @0xdiid Think they've always allowed it.

A while ago they said they would stop allowing it, but then reversed that decision a few weeks later.

[wireless @wireless_dev]

@thdxr are u gonna open source the anomaly blog post.

told my agent to make a PR explainer in the style of that post, didnt realize it was closed source. still came ok though.

[wireless @wireless_dev]

we've grappled with this for a while (obv smaller scale). how do you do this, even when you get large enough? your traffic will still be uneven over the day / week.

i'm guessing the labs get away with this because they can "monetize" those idle gpus by using the compute for research, which is less time sensitive.

[wireless @wireless_dev]

@Tommyziv7 @colinhacks Maybe only supported in contexts with eval / Function constructor enabled?

[wireless @wireless_dev]

@GregKara6 @mattzcarey Instead of doing 10 steps chained together (each one having all of the previous ones in its context), delegate each step to a subagent, passing along only the context it needs.

[wireless @wireless_dev]

@thdxr technically a DAG but you get the idea

[wireless @wireless_dev]

@thdxr yea I think of our app roughly structured as a tree

frontend files are the leaves at the bottom - they depend on other pieces of logic, but no logic depends on them

slop is more acceptable at the leaves than the stuff higher up

[wireless @wireless_dev]

@thdxr We run fully on Cloudflare w/ SST and no wrangler. What do you need wrangler for specifically?

[wireless @wireless_dev]

@trikcode You can tell GPT to read claude's context. It's all stored locally by claude code.

[wireless @wireless_dev]

@marty_kausas @Dee_Tay_ler @conductor_build Conductor is still the answer

[wireless @wireless_dev]

I'm skeptical this is the best approach

A lot of times, you don't know actually know how hard a problem is until you actually start working on it.

The real solution imo: let the big model the call the small model as a subagent. It can assess when to offload work on its own.

[Brian Armstrong @brian_armstrong]

How to keep AI spend flat while token usage grows exponentially: Not with friction and spend alerts. With better defaults, routing, and caching.

Better Defaults (not Usage Caps) – Engineers can choose any model they want, but defaults matter. We’re experimenting with defaulting to open weight models like GLM 5.2 and Kimi 2.7 through our LLM gateway, while still encouraging engineers to choose the right model for the task. 91% of our employees were never hitting their usage caps, so instead of lowering caps and driving up alerts, we're moving to cheaper defaults. Note that code reviews use a diversity of models, so they can check each other's work.

Better Routing – In our custom harnesses, we preprocess prompts and route to the best model for the job, considering cache hits and model pricing. For instance, you may want a frontier model for planning, but not for execution where they can be overkill. Ultimately, humans shouldn't be choosing models - AI can automate this task.

Better Caching – Cache misses are the easiest way to drive your cost up. All of our requests are cache aware, so we’re reusing a warm cache wherever possible. For example, our cache hit rate went from 5% → 60% in LibreChat once properly implemented.

Keep Context Lean – Start fresh sessions when switching tasks. Scope file context narrowly. Disconnect unused tools. Don't just compact. The goal isn't fewer tokens used, it's fewer tokens wasted.

Better Visibility – Our engineers can use as many tokens as they want, from whatever model they want, but we’ve made usage visible – and the more you spend on AI, the more impact we expect.

The goal isn't to suppress usage. It's to build the infrastructure that makes exponential growth sustainable.

Putting this into practice has cut our AI spend nearly in half, while our token usage continues to grow.