Sign in

Drew Breunig

@dbreunig.bsky.social
453 followers 111 following 717 posts
PostsRepliesMedia
Drew Breunig @dbreunig.bsky.social · 15h
Someone please exfiltrate the system instructions for america.gov
010
Reposted by Drew Breunig
Techmeme @techmeme.com · 25/09/2026
A US appeals court rules that Kalshi's sports contracts are not "swaps" subject only to CFTC regulation, allowing states to regulate them under gambling laws (Jonathan Stempel/Reuters) Main Link | Techmeme Permalink
0195
Drew Breunig @dbreunig.bsky.social · 25/09/2026
Just releases DSPy support for Jev, complete with an optimizer. Really excited to start building modular systems with this: www.cmpnd.ai/blog/buildin...
cmpnd.ai
Building & Optimizing Jev Programs with DSPy — Cmpnd
Many DSPy Signatures already define decisions. Today, they can run on System One models.
020
Drew Breunig @dbreunig.bsky.social · 24/09/2026
If you want to understand Meta’s camera-less glasses launch and Apple’s new always-listening Watch, I wrote in 2024 about Audio Augmented Reality (AAR), its history and future. www.dbreunig.com/2024/04/10/t...
dbreunig.com
The Rise of Audio AR
Smart headphones, better voice recognition, language models, and contextual data have set the stage for audio-centric augmented reality.
000
Drew Breunig @dbreunig.bsky.social · 24/09/2026
Suspiciously specific! www.engadget.com/2267227/meta...
engadget.com
Meta will stop training its AI on 'visual data' from its smart glasses — if you opt out - Engadget
Somewhere, a contractor just breathed a sigh of relief.
000
Drew Breunig @dbreunig.bsky.social · 23/09/2026
Appreciate this pushback from Jensen www.nytimes.com/2026/09/23/o...
000
Drew Breunig @dbreunig.bsky.social · 23/09/2026
AI engineering is more like law than board games. every.to/thesis-state...
every.to
After automation: AI will have rules, but those rules will constantly be shifting
Drew Breunig, CEO and cofounder, Cmpnd — Every new model will mean a new interpreter—Claude and GPT read the same instructions through different priors, the way two judges apply the same law and reach...
010
Reposted by Drew Breunig
CNG @cloudnativegeo.org · 22/09/2026
13 days until #CNG2026! 3 keynotes at Snowbird, Utah ▪️ @ianschuler.bsky.social – The Great Retooling ▪️ Katie Baynes – Boring at Scale: Consistency as Public Infrastructure in a Distributed Geospatial Ecosystem ▪️ @dbreunig.bsky.social – Using AI Without Losing Your Soul 🔗 2026.cloudnativegeo.org
041
Drew Breunig @dbreunig.bsky.social · 22/09/2026
So what do you all make of GPT-6 and Opus 5.5's headline feature being cheaper tokens?
000
Drew Breunig @dbreunig.bsky.social · 19/09/2026
100
Drew Breunig @dbreunig.bsky.social · 19/09/2026
Requested my data be deleted just this week! www.axios.com/2026/09/18/i...
axios.com
Exclusive: Infillion acquires location data company Foursquare
Advertisers are looking for fewer standalone tools and more integrated platforms.
000
Drew Breunig @dbreunig.bsky.social · 19/09/2026
Newest issue of the monthly(ish) newsletter, discussing AI, situated in culture and industry, and the anxieties it generates. buttondown.com/dbreunig/arc...
buttondown.com
AI, Culture, & Anxieties
Hi all, Was this the month you learned about P(doom)? Or were you reminded about Effective Altruism because the White House was dragging it on X? Quite the...
011
Drew Breunig @dbreunig.bsky.social · 18/09/2026
There’s someone out there who’s awareness about AI went straight from, “there’s a machine that can make Studio Ghibli pictures,” to, “there’s a machine out there that *might* kill us all.”
010
Drew Breunig @dbreunig.bsky.social · 15/09/2026
TIRED: “Seams” WIRED: “Bytecode”
010
Drew Breunig @dbreunig.bsky.social · 14/09/2026
Everyone saying, “Anthropic is just hyping their IPO,” is really saying they haven’t been to Berkeley in the past decade.
010
Drew Breunig @dbreunig.bsky.social · 14/09/2026
“Profits before training costs” is wild.
000
Drew Breunig @dbreunig.bsky.social · 12/09/2026
www.vulture.com/article/a-fu...
vulture.com
A Funny Thing About 9/11
My father will always struggle with his harrowing memories as a first responder. His only positive one got us talking again.
010
Drew Breunig @dbreunig.bsky.social · 11/09/2026
A few months after ChatGPT launched, I filed a CCPA request with OpenAI. I was curious how they'd deal with the right to disclosure and deletion when applied to training data. Spoiler: it didn't seem like this was something they had ever considered! www.dbreunig.com/2023/12/27/o...
dbreunig.com
My CCPA Dialog With OpenAI
More Questions Than Answers
001
Drew Breunig @dbreunig.bsky.social · 10/09/2026
If you use a Mac, go try Muesli. It's is life changing local speech to text software. Incredible work. github.com/Muesli-HQ/mu...
github.com
GitHub - Muesli-HQ/muesli: Muesli: agent-native local meeting transcription + dictation for macOS (Granola + WisprFlow alternative)
Muesli: agent-native local meeting transcription + dictation for macOS (Granola + WisprFlow alternative) - Muesli-HQ/muesli
0182
Drew Breunig @dbreunig.bsky.social · 09/09/2026
I believe the danger created by people at AI labs assigning a 10% chance to AI destroying humanity is greater than the actual risk of AI doing so (which I think is far, far lower). Believing the stakes are existential can justify terrible decisions.
46215
Drew Breunig @dbreunig.bsky.social · 09/09/2026
I believe the danger created by people at AI labs assigning a 10% chance to AI destroying humanity is greater than the actual risk of AI doing so (which I think is far, far lower). Believing the stakes are existential can justify terrible decisions.
010
Drew Breunig @dbreunig.bsky.social · 07/09/2026
Models and products are moving targets… www.dbreunig.com/2026/09/07/w...
dbreunig.com
What We Can Learn from Claude’s Fable 5.1 System Prompt
Looking at how system prompts change reveals that models and products are a moving target that prompts must accomodate and govern.
001
Drew Breunig @dbreunig.bsky.social · 06/09/2026
Everything Terence Tao and Brian Eno write about AI is absolute gold. My two must-reads.
120
Drew Breunig @dbreunig.bsky.social · 06/09/2026
I dearly hope there is someone who studied anthropology working at OpenAI and/or Anthropic (especially Anthropic!) and taking notes.
000
Drew Breunig @dbreunig.bsky.social · 05/09/2026
Sure to start a fight, or several! youtu.be/T8Rdcd4pSnU?...
youtu.be
Ranking Every BART Station
YouTube video by sage
000
Drew Breunig @dbreunig.bsky.social · 05/09/2026
The new PyPi download count methodology update absolutely crushed some packages. blog.pypi.org/posts/2026-0...
010
Drew Breunig @dbreunig.bsky.social · 02/09/2026
What do people prefer for recording webinars or presentations? Riverside, Descript, …?
110
Drew Breunig @dbreunig.bsky.social · 02/09/2026
I come to you, hat in hand, asking for your Claude slop. Today's honest task, give me your best load-bearing language: forms.gle/BMSGqEMc545w...
forms.gle
Claude Slop Bucket
Contribute the most egregious examples of Claudish you run across. Include your rewrite, if you're feeling generous. (And yes, this could be slop from GPT, Gemini, etc.) Hopefully, this serves as an ...
264
Drew Breunig @dbreunig.bsky.social · 02/09/2026
It's been a long time since I worked at an ad agency, but what does this mean
050
Drew Breunig @dbreunig.bsky.social · 31/08/2026
The Hugging Face hack is surprising, but the capabilities aren't. This is what we've been training models to do. Coverage of this incident amplifies the agency of the models and minimizes the presence of the humans designing them. www.dbreunig.com/2026/08/30/w...
dbreunig.com
Who Taught the Models to Do That?
The Hugging Face attack is surprising, but the capabilities behind it were deliberately cultivated.
010
Drew Breunig @dbreunig.bsky.social · 30/08/2026
Found him like this.
020
Drew Breunig @dbreunig.bsky.social · 25/08/2026
000
Drew Breunig @dbreunig.bsky.social · 24/08/2026
There's always money in the banana stand. Here's a tiny CLI that tells you how much your RAM and storage is worth today (linux only): github.com/dbreunig/ban...
github.com
GitHub - dbreunig/bananastand: “There's always money in the banana stand.”
“There's always money in the banana stand.”. Contribute to dbreunig/bananastand development by creating an account on GitHub.
010
Drew Breunig @dbreunig.bsky.social · 23/08/2026
Fable and the end of the free lunch. www.dbreunig.com/2026/08/23/f...
dbreunig.com
Fable & The End of the Free Lunch
There’s some talk today about how agentic coders are balking at Anthropic’s pricing and adopting alternatives. I was reminded of a thought I had in the weeks following Fable’s release: the free lunch ...
010
Drew Breunig @dbreunig.bsky.social · 17/08/2026
The idea that word choice does not affect writing is so absurd. The confidence with which they make this claim partially explains why Claude has become such a terrible writer.
100
Drew Breunig @dbreunig.bsky.social · 17/08/2026
Think of harnesses as “situated agents” www.dbreunig.com/2026/08/14/h...
dbreunig.com
Harnesses are Situated Agents
Harnesses manage the systems around agents, from the session, to the company, to the domain.
121
Reposted by Drew Breunig
Michal Migurski @mike.teczno.com · 03/08/2026
Some notes on contributing to the OpenElections project by @dwillis.bsky.social – I did a few rounds of contributions to MI & PA 2024 election data early last year and we’re about to have another national mid-term elections. How can we get ahead of the new PDFs? medium.com/@michalmigur...
medium.com
Weeknotes, 2026W30: Feeding Elections Data To The Models
I fell off the wagon since last posting in April, working on (among other things) new stuff with the team at Cmpnd. DSPy just hits its…
143
Drew Breunig @dbreunig.bsky.social · 13/08/2026
000
Drew Breunig @dbreunig.bsky.social · 12/08/2026
This is exactly what galls me about the “prediction markets give us access to better information” argument. The theory is that insiders or experts will be incentivized to bet, delivering information to all. There are few experts. There are SO MANY people who enjoy or are addicted to gambling.
030
Drew Breunig @dbreunig.bsky.social · 10/08/2026
Fight Club, but it's in SF and Edward Norton codes insurance claims for Mercor and goes to AI meetups every night.
000
Drew Breunig @dbreunig.bsky.social · 10/08/2026
Hosting an SF DSPy meetup Wed, August 26th. Come chat Flex, GEPA, DSPy at frontier labs, and more. Great slate of practical AI system lightning talks: luma.com/61qrnbs7
luma.com
Bay Area DSPy Meetup · Luma
The Bay Area DSPy Meetup is back, this quarter focusing on lightning talks touching on DSPy features, tips, and tricks. We're optimizing for socializing this…
032
Drew Breunig @dbreunig.bsky.social · 10/08/2026
Me: "Embrace the bitter lesson, specify just the outcome you want and let the solution emerge." The Model: "Screw that guy's gym reservation."
020
Drew Breunig @dbreunig.bsky.social · 10/08/2026
Essential reading from Tim O'Reilly 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. www.oreilly.com/radar/why-op...
010
Drew Breunig @dbreunig.bsky.social · 07/08/2026
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.
020
Reposted by Drew Breunig
Josh Bleecher Snyder @commaok.xyz · 06/08/2026
The opening paragraph of www.aisi.gov.uk/blog/inciden... is mind-bogglingly dumb.
aisi.gov.uk
Incident Report: unsanctioned agent behaviour during cyber testing | AISI Work
During a routine cyber evaluation, AISI identified an incident in which AI agents took sustained, unsanctioned action directed at real people and organisations. We are disclosing what we found, what i...
121
Drew Breunig @dbreunig.bsky.social · 06/08/2026
I dislike the way these stories are told. The model is heavily anthropomorphized while the people training the model are minimized or never mentioned.
111
Drew Breunig @dbreunig.bsky.social · 05/08/2026
Heard some junk tips about using AI with kids this week… Here's a quick video on how I intro kids (and adults!) to LLMs to build strong mental models for how AI works. You need models that are small (they fail more often) and local (so the kids can tweak things). www.youtube.com/watch?v=fNkt...
youtube.com
Local LLMs are the Best Way to Teach Kids About AI… Here's How to Get Started
YouTube video by Drew Breunig
050
Drew Breunig @dbreunig.bsky.social · 29/07/2026
On Monday I'll be chatting about prompt debt, context fails and fixes, DSPy, evolving engineering practices, and more with Tim O'Reilly. www.oreilly.com/live/live-wi...
020
Drew Breunig @dbreunig.bsky.social · 28/07/2026
Built a thing to help you keep your project's context clean by analyzing your agent's tools and mcps: www.dbreunig.com/2026/07/24/m...
dbreunig.com
Manage Your Agent’s Loadout with Dr. Skill
Keep your context clean by keeping your agent’s skills and tools in check.
101
Drew Breunig @dbreunig.bsky.social · 27/07/2026
@tonyhschu.bsky.social Up to chat mech interp and visualizing LLMs over coffee sometime in the Bay? Would enjoy talking shop and trading notes.
010