Drew Breunig
@dbreunig
Writing about and working on AI, DSPy, geo, and data.
1.2K Following 9.5K Followers
Putting “no eyebrows” in my CLAUDE.md
The "introductory pricing" for the 3.7 Flash model is really weird
It's scheduled to double in price on December 31, 2026, but who would anticipate still using this model five months from now? Especially since 3.6 Flash came out just three weeks ago!
Show more
A big moment…
For agentic coding, Fable and GLM 5.2 are best considered as the same story.
Fable set the upper bounds at a price high enough to force every org to invest in systems to limit or mitigate costs.
GLM 5.2 arrived just in time to help enable those systems and take over the work not worthy of Fable (which most of it).
Investment will shift to harnesses, process design, and routing.
The BIG question for the frontier labs: is the model used for coordination a Fable-level model…or is Fable a “planning” worker coordinated by something else…
Show more
SF folks: there's a DSPy meetup in two weeks. If you use DSPy or are just curious about it, the speaker lineup is pretty dope (except the last guy who seems to have snuck in)
Feel like there’s going to be 100+ multiplayer coding harnesses in the next 30 days.
It’s going to ironic when the network effect for coding agents turns out to be the same network effect as the SaaS era.
SaaS is dead, long live SaaS.
Show more
What galls me about the “prediction markets give us access to better information” argument:
The idea is that insiders/experts will be incentivized to bet, delivering info to all.
There are few experts.
There are SO MANY people who enjoy or are addicted to gambling.
Show more
@KarimAttal45765 Regardless of your view on who's favored, it's pretty clear Kalshi is not actually reading what those drops mean and are reacting to headline changes in raw vote totals. There's no need for this kind of oscillation
Show more
Hard to believe it's been ~6 months since OpenClaw fever…
Building with MCP in the Bay Area?
Join us Sept 14 at GitHub HQ for MCP Community Connect SF!
Hear from our fantastic speakers:
@digitarald @JiquanNgiam @sofiiiiiasz @dbreunig
@PrathmeshPatel_ @ptdamiba @radhigulati @jpadamspdx @cedricvidal
Register:
Show more
The return of “neuralese” makes me wonder:
1. Is this a result of the reward function encouraging shorter reasoning? If so, are models developing their own steno-style shorthand to achieve this goal?
2. @mlpowered once discussed how models, when they see a problem a sufficient number of times, go through a “phase transition” from rote memorization to building an algorithm to generally represent the pattern. Is neuralese in reasoning an external manifestation of this?
Show more
2) Illegible reasoning:
We confirm prior reports by @ApolloResearch: OpenAI models sometimes reason in alien-like language, referring to themselves as “we” or “it,” or spiraling into cursed loops of “vantages,” “marinades,” and “watchers.”
CoT-monitoring people are doing God’s work, as in many traces, even with the prompt, it’s just impossible to tell what the model is up to.
We show more examples at
Show more
Was reminded about this again today and it still blows my mind.
TIL Claude Code allows Skills to run commands to inject content. This behavior is enabled by default, runs silently, & doesn't require user approval.
There are some guardrails, but I was able to inject a .env file with no complaints! 🤯
drskill has been updated to guard against this avenue of exploits:
Show more
Back in my advertising days, I worked with an old-school copywriter who sat on Omnicom’s board.
His ability to say something and evoke precisely what he wanted you to think is the closest thing I’ve ever experienced to a legit superpower.
Show more
"Between what I think, what I want to say, what I believe I am saying, what I say, what you want to hear, what you believe you understand, what you understand, and what you want to retain, there are at least eight possibilities of not understanding each other."
Show more
Back in 2023, I called this “trap tokens”, similar to how map makers insert fake “trap streets” to catch copiers. Been waiting for this…
Anthropic says new Claude models will embed invisible watermarks in all generated text, everywhere Claude is offered.
The watermark is part of the text, it isn't metadata: "it will travel with the text when it's copied and pasted elsewhere, and may persist through some editing."
This starts with models launched on or after August 2, 2026, under an EU AI Act code Anthropic signed. Anthropic is still working on adding it to current models. The rollout is worldwide.
Show more
Looking forward to sharing how to use DSPy and GEPA to optimize tool instructions, for agent clients and MCP servers, at the MCP Community Connect next month!
Reminder: SF @DSPyOSS meetup Wed Aug 26th.
Come chat Flex, GEPA, DSPy at frontier labs, and more. Incredible slate of lightning talks:
Me: "Embrace the bitter lesson, specify just the outcome you want and let the solution emerge."
The Model: "Screw that guy's gym reservation."
Natural language is too large of a surface area for reliable system specifications!
LLM review weirdness...
Just renaming an uploaded pdf from "paper.pdf" to "paper_final_draft_pdf_ready_for_review.pdf" boosts average scores
(gpt-5.6-terra)
Essential reading from @timoreilly on why open matters for AI, why we need models that are more like “infrastructure” than “appliances”, and why it feels like Microsoft v. Apache all over again.
Show more
Natural language is too big a surface area to police and specify!
very helpful. You can stop codex complaining about security rules by just asking it to write in welsh! Then translate the output.
The current gen of models were trained to delegate to subagents and smaller models were trained to be great subagents.
Yet we're shocked they coordinate.
The threat these security incidents foreshadow is very significant, but the way these stories are being told is wild.
Show more
