Co-Founder & CEO at @AstridStyleAI
New York, NY
Joined June 2008
- Tweets5K
- Following215
- Followers3.2K
- Likes2.7K
Kyle Rush retweeted
sonnet 5.5 is an absolute token-hungry monster
5.5 max — 193k output tokens/task
opus 5.5 max — 119k
sonnet 5 xhigh — 65k
gpt-6 sol max — 31k
gpt-6 astra max — 27k
using 62% more output tokens than opus 5.5 and over 6x more than gpt-6 sol
Kyle Rush retweeted
"Hint is very able to look at that three d model, look at all the situations, and say, you know, Kyle, your the back of your house, that deck basically gets no sun exposure every day of the year."
Kyle Rush, Co-Founder and CTO of Hint, says its 3D home model finds hidden causes and surfaces problems homeowners do not know how to ask about.
"How do they even know to ask that wood rot can be solved? Like, how do they even know what wood rot is?"
@kylerush
Kyle Rush retweeted
"everything is just wrong"
Kyle Rush, Co-Founder and CTO of Hint, says generic AI gives wrong home advice when it misses local jurisdiction and property context.
"whenever I talk to any of these AIs, even my work Claude that I work on, you know, on Hint, it still thinks that I'm in town of, Bedford."
"Anytime I ask about taxes or regulations or how many chickens we can have on our property, it's all just wrong."
"until there is some way that, like, Claude figures out how to correct that problem, I think you're just gonna be getting a subpar experience."
@kylerush
Kyle Rush retweeted
We’ve identified industrial-scale distillation attacks on our models by DeepSeek, Moonshot AI, and MiniMax.
These labs created over 24,000 fraudulent accounts and generated over 16 million exchanges with Claude, extracting its capabilities to train and improve their own models.
Yes.
Never run queries without a timeout.
A single rogue query or load spike can lead to stalls or app downtime without them. Let's think through why.
Consider with case where a long-running query gets inadvertently introduced into our app. The database typically processes only short-lived queries (10ms or less) and suddenly we have a new query that takes 1000ms per execution.
Not only will this hog resources, but will shoot up the number of simultaneous transactions. We'll either hit connection limits, txn limits, or use 100% of the database cpu / iops.
Now consider the same scenario, but we place a 250ms timeout on every transaction (database-side), with exponential backoff retry logic + jitter (app-side). We've now limited the blast radius of any single query. Long-running queries will get killed, and the backoff logic minimizes the risk of thundering herds.
By monitoring these timeouts, we can quickly identify the problematic query and roll back the change.
Kyle Rush retweeted
We're at an interesting stage of the evolution where the models are 1000x faster than I am at coding and 10x slower at using a web browser.
I went all-in on agentic coding. Told my team I'd be in no meetings for a full week, declined everything, and just did this. Eight Cursor agents running in parallel, each in its own isolated devcontainer. 46 pull requests merged. ~30,000 lines changed. Over $1,000 in inference costs. I think only as of last month — because of Opus 4.5 — was all of this possible. Now we're on Opus 4.6.
I wrote about everything — what worked, what broke, and what I learned. My six takeaways:
1. Engineers are becoming engineering leaders of elastic teams. The job is orchestration, not typing.
2. Tooling speed suddenly matters a lot. If feedback loops aren't fast, parallelism collapses.
3. You have to trust agents, but strong observability and fast recovery must be part of the system.
4. The real work is planning and continuously improving the rules and constraints agents operate under.
5. If your inference bill is low, you're probably not using AI. Optimize for results, not cost.
6. The agent's output will be as good as the plan you build with it.
kylerush.org/posts/opus-4-5-…
Kyle Rush retweeted
Every test framework I've ever used that chooses to captures logs (stderr, console.log, etc.) has managed to make my life harder by doing so. Please, if you're making a testing framework, do not capture logs.
Kyle Rush retweeted
AI users are working longer because, for the first time, our tools can finally keep pace with our imagination.
But speed doesn't matter if you lack a destination.
When output is cheap, your greatest leverage isn't how much you build, but the taste to know what is worth building - and the rest required to see the difference.
😆
Python is the language of AI. By all conventional measures, it shouldn't be.
This unserious, slow, and unsafe language is the best worst thing to happen to programming – and we love it.
@PrefectIO’s @AAAzzam and our Head of AI @BEBischof discuss why “the worst language” became the foundation of the AI stack: theoryvc.com/blog-posts/the-…
Add one in the Team Opus column. If Codex actually learned and became a better engineer as I coded with it and guided it, I might change my mind. But it doesn’t do that. No coding agent does. It is not a good experience to be telling an AI the same thing over and over.
Just tried Opus 4.6 Thinking in Cursor and boy was it super slow. I’ve ever seen a model think that much. Totally unusable. I’m hoping this is just because their server is overloaded because of the launch.
