Drew Breunig
16K posts
Drew Breunig
@dbreunig
Writing about and working on AI, DSPy, geo, and data.
Bay Area Katılım Mart 2008
1.2K Takip Edilen9.3K Takipçiler
How I imagine OpenAI created the training data for their current image model:
English
@raveeshbhalla In a very small sample set I have been enjoying the quality, speed, and cost of Kimi. It feels at least as good as large GPTs.
English
@dbreunig Need to explore the newer & gemini models a bit more but generally "OpenAI good, Anthropic bad" x.com/raveeshbhalla/…
Raveesh 折図@raveeshbhalla
GPT 5.4 mini as the student model + GPT 5.5 as the teacher model for GEPA runs is the most powerful combo I've seen. The latter is so incredibly good at steering the former.
English
@swyx True, but the tip of the iceberg.
I recently heard from someone in a large company whose coding agent was configured to load ALL skills any employee had pushed into shared repo.
Unless the user knew about this, 300+ skills were injected on every call.
English
cosign. models have overtuned to this now and do not realize when the agentsmd is out of date and should be changed/ignored.
last night i goaled 5.6 sol to complete a 5 stage task and woke up to find it was still stuck on stage 0. it took a while to read the transcript back a few hours to realize at some point some agent had committed “stage 0 is the target dont do anything else” so poor sol spent 8 hours only refining and verifying stage 0 because /goal would not let it stop and agentsmd would not let it proceed.
if you dont know whats in your agentsmd before you fire off each task, it is an indirect prompt injection you perform on yourself. /plan, /goal, /skill, or nothing at all.
Ryan Dahl@rough__sea
AGENTS.md/CLAUDE.md is largely an anti-pattern
English
My first internship after college was under someone who thankfully made it his mission to get the academic writing out of my system. I vividly remember the constant, "WHY DO I CARE?" notes that were written over every draft.
It taught me two essential lessons for writing:
1. Attention economics are real. If you list 10 things, your audience will remember only 2-3 (if you're lucky!) and you don't get to pick which ones they do remember. So hone your argument, cut down your core points, and force the choice by giving them fewer options.
2. The takeaway you're trying to convey is too important to leave unsaid. Be sure you tell them _why they should care_. (You have to earn this obviously, but _do not leave it unsaid_.)
English
i've been letting fable run experiments (mostly autonomously) to test a few ideas we had about jlens
when does the structure form in training? can you transfer between models? how does it scale? how far does a nudge travel? what does K2 jlens look like?
eliebak.com/viz/jspace-ope…
elie@eliebakouch
computed the similarity (CKA) on the J-lens geometry of every layer inside and across 38 open models. the patterns are weirdly universal: same depth layout, same organization at the same relative depth, even between unrelated families like llama and olmo eliebak.com/viz/jspace-open
English
Perhaps your audience reads this way, but would encourage you to check a couple boxes to potentially reach a wider audience:
- Set up the question clearly and why it matters
- Enumerate your findings simply
- Then put your finger on the "so what" so people can connect the dots.
I like the 'prompt aside' at the top that lets people peak behind the curtains of how you got here, but I (and I'm sure others) want YOUR interpretation of what you found and what its interesting before wading through the exploration.
English
At the moment, model builders are trading diversity for reliability.
It is the human's (and a lesser extent, the harness') job to injecting diversity, creativity, and push the model out of distribution.
English
@dbreunig Is it a reward hack? Like it's one way to sound smart without more substance.
Brooklyn, NY 🇺🇸 English
What is your theory for why LLMs lean so consistently on emdashes?
My top two:
1. Much of their training data is transcribed speech and humans speak in long, compound sentences.
2. There's a reward function, somewhere, that encourages the model to use fewer sentences.
English
“If Anthropic’s technology is as powerful as they say it is – so powerful that open models like it should likely be banned – then they should be able to secure their API.”
Nathan Lambert@natolambert
The open model community is extremely unprepared for when a model gets stuck in the undefined white house licensing regime - and it could permanently knee cap the open model economy within 6 months. Why this'll happen and what we can do: interconnects.ai/p/6-months-to-…
English
Where do people show off their information design work these days?
English
@dabblingfrancis Claude and GPT out here using it multiple times every paragraph.
English
“We're going to upload all your code, even if you're not working on it, unless you specifically opt out,” is a wild position.
SpaceXAI@SpaceXAI
We care deeply about your privacy and respect customer choice. For teams using zero data retention, no trace and code data is ever retained. All API key use of Grok Build also respects ZDR. If ZDR is disabled, the /privacy command is available in the CLI to disable data retention, which also deletes previously synced data. Run the /privacy command to view or change your settings at any time.
English
@bradenjhancock There are system prompts and/or training that encourages the models to cite specific figures and ask for a close, and it takes so much prompting to get around it.
English
@bradenjhancock The number one Claude tell is repeating the same figures on the page several times. Once in the headline, once in body copy, once in the big figure boxes it loves so much. I see it everywhere.
x.com/dbreunig/statu…
Drew Breunig@dbreunig
This feels right. Slop occurs because people think AI's first shot is sufficient. It takes many turns of honing to exceed the communicative quality of your initial input. If you aren't going to spend the time and turns, send your original prompt as the email.
English
There's a very noticeable smell to AI-generated pitch decks, and these decks have become noticeably more prevalent in the past few months.
- Design: You can tell it to change the fonts or colors to match your brand, but unless you're really intentional, you're going to get the same go-to templates, box styles, font ratios, etc. as everyone else.
- Information density: Masterful storytellers make one point per slide and vary slide density, similar to how masterful writers make their content interesting to read with a short sentence, then a long compound one, then a medium one. AI-generated decks truck along with slides that are ~75% full. There's some whitespace and margins, but way too many visual components and wasted words.
- Tone: It's hard to put my finger on it, but the really great founders and decks I've seen don't feel like they're selling something. They feel like they're stating how the world is and clearly will be and you can come along for the ride as an investor if you'd like and ask nicely. AI-generated decks are clearly selling something.
English
Every time we get a new technology, we shoehorn its capabilities into a form factor that fits our infrastructure, culture, and habits. We use it in a way we currently know. This is never the final form.
Here's one of the first automobiles:
Simon Willison@simonw
The idea of "AI employees" feels so short-sighted to me - both disrespectful to humans and a complete misunderstanding of what these tools can do and how to best put them to work You may as well start adding Excel spreadsheets to your org chart
English
Drew Breunig retweetledi
The idea of "AI employees" feels so short-sighted to me - both disrespectful to humans and a complete misunderstanding of what these tools can do and how to best put them to work
You may as well start adding Excel spreadsheets to your org chart
English