@ClementDelangue GPT-image-2 diagrams are a step change in showing off data for social media. Created a few internally at work that were big hits.
English
David Rose 🫧
1K posts
David Rose 🫧
@drose101
Randomness and dynamical systems. ML engineer.
New York, NY Katılım Eylül 2010
522 Takip Edilen354 Takipçiler
Going through my old blog posts and converting them to interactive demos. This is so much better than markdown or a PDF.
Here we are visualizing the token generation logits between two quite similar prompts and trying to understand where they diverge.
English
Why is Claude asking me so many follow up questions this week. It’s like each task I’m given a multiple choice survey to fill out before starting.
English
The German Army is testing a 10 kW laser weapon system known as JUPITER (Joint Universal Platform for Laser Integration, Test and Evaluation in Realtime), mounted on the Boxer.
The system was demonstrated during trials of future combat concepts in Münster on April 30, 2026.
JUPITER is a German–Dutch collaborative program built around a modular mission package integrated into the Boxer platform. This design allows laser weapon technology to be quickly installed, removed, or upgraded within a standardized vehicle module, making the system flexible and scalable for different mission profiles.
1/
𝔗𝔥𝔢 𝕯𝔢𝔞𝔡 𝕯𝔦𝔰𝔱𝔯𝔦𝔠𝔱△ 🇬🇪🇺🇦🇺🇲🇬🇷@TheDeadDistrict
In Münster, the German Army and industry are demonstrating what the future of ground combat could look like during an experimental exercise. Unmanned and manned ground and aerial systems are taking part in the demonstration. Let's break down what they showcased: Boxer Block II
English
Big day for normies around the world.
OpenAI@OpenAI
GPT-5.5 Instant is starting to roll out in ChatGPT. It’s a big upgrade, giving you smarter, clearer, and more personalized answers in a warmer, more natural tone. And it's also more concise, which we heard you wanted. We think you'll love chatting with it.
English
@ericmitchellai If it stops emoji-dumping like 5.3-instant I will be happy. Though on the flip side having such a tell was useful in many situations.
English
Excited that we're updating the default model in ChatGPT today!
5.5 instant is a substantial improvement in intelligence, image perception, and factuality.
It also updates the writing style to be a bit plainer and more straightforward.
What was on your wishlist?
OpenAI@OpenAI
GPT-5.5 Instant is starting to roll out in ChatGPT. It’s a big upgrade, giving you smarter, clearer, and more personalized answers in a warmer, more natural tone. And it's also more concise, which we heard you wanted. We think you'll love chatting with it.
English
LLM writing in contexts outside of code is still so terrible. And it’s all bad in the same way with the same tells, which is somewhat surprising to me.
Are people fine tuning on their own writing? Is promoting enough? Posting this so I remember to solve this tomorrow.
English
Cooking up some fun discussions for a talk next week.
Finding chaos in the model
English
After spending the past few years traveling and living out of hotels and airbnbs across the world, I can confidently declare that USA grocery stores are S-tier with no competition.
No I don’t want to spend 2 hours visiting a different market for each of my meats, bread, and veggies. It’s a bit fun and cute at first but efficiency wins out in the end.
English
i'm a few days late to realizing this but:
wow, opus 4.7 is god awful
like so, so bad
it's making mistakes on things i'd expect gpt-4o to handle cleanly
there's got to be some explanation, right?
English
Coming across this book in my Airbnb is like rekindling with an old love. It was the first time I connected with a quantitative subject, having struggled through most of mathematics in my earlier schooling.
This dude seemingly has two passions according to his home library: finance and communist literature.
English
iOS live activity and lock screen usage are underutilized. Along with Dynamic Island updates. Duolingo uses it somewhat annoyingly and uber uses it nicely, But it’s a sort of race to the bottom, though the race hasn’t started yet. Maybe Apple sort of pushes back against the behavior.
I want to hook up my agents to sync realtime to lock screen. It’s the final evolution of coding on-the-go.
I tried PWA, it feels jank. Tried home screen widget, too many restrictions. But live activities kit is sort of a Wild West. You have an 8 hour limit but can simply refresh it. ActivityKit is powerful and combined with APN you can get an almost app-like realtime vibe.
English
@gregpr07 Got it! No hate I just happened to be doing some eval on this new tool at work a couple days ago with my agent and it mentioned the task assignment. Was in the back of mind while scrolling tweets just now.
English
> flexes star counts
> Browser Harness onboarding literally has the agent star its own repo
Gregor Zunic@gregpr07
I love open source. You can provide so much value to the world. Do competitors steal our stuff immediately? Yes Do I really care? No We raised 17M and burned almost nothing so we can keep doing this until everyone else runs out of money.
English
Nvm. At current prices Deepseek v4 flash has replaced grok as my grunt worker llm.
EXCEPT it doesn’t support vision input?? “Support incoming” from what I can see.
English
Everyone makes fun of the @xai grok models but grok-4.1-fast is legit a solid choice for general usage tasks.
- 2 million token context
- $0.20 / $0.50 per M tokens! That’s wild cheap lol
- 130 tokens/second
I ship and archive every single AI agent message and tool call into a centralized db with summarizations and classifications after *every turn* and this is all done via grok-4.1-fast for maybe a few dollars a day.
Also useful for analyzing video and images in an ongoing process due to low cost, using it for video feeds and alerting on security cameras.
English
code review is now just my teammates having their AI comment on my PR while I just task my AI to respond to their AI. And so on and so forth.
English
We should just let the agents pick and choose what to remember or forget. Each round they pick and choose from some index. Like how Claude manages an ongoing task list, it knows best.
English