@carlosbedia @mattpocockuk Makes sense. tbh my wording was a bit harsh
Input: you have full control
Output: depends on the model, you can just “cut off” at the limit. It tries to respect your wishes but it’s mostly best guess afak
English
Felix Dolderer
254 posts
Felix Dolderer
@felix_dolderer
Create the future you want to live in.
Near the beach, CET Katılım Nisan 2026
169 Takip Edilen4 Takipçiler
@felix_dolderer @mattpocockuk My point was to ask since I did not know. It seems some still have 'max_tokens' or 'max_completion_tokens'. Grok sends the last message with tag "Length", OpenAI for runs lifecycle, Anthropic as a StopReason. So it seems you have some control in Grok
English
The "X technique reduces tokens by Y%" fad is so old
Can't believe people get taken in by this
Charly Wargnier@DataChaz
UP TO 95% TOKEN REDUCTION WITH ZERO CODE CHANGES A Netflix engineer just open-sourced Headroom, and it’s one of the smartest ways I’ve seen to cut LLM costs. It wraps Cursor or Claude in a local proxy to compress your payload before it hits the LLM: → Intelligently shrinks logs, JSON, and code → Perfectly preserves logic accuracy → Keeps 100% of your data local → Stops Opus-tier models from wasting tokens on boilerplate It already crossed 35K stars, which says a lot. 100% free and open-source. repo in 🧵↓
English
@JaimeOrtega @mattpocockuk Some compressions work - I have even implementedg one of these myself a while back.
Say you have 500 tests and each passing test logs 3 lines: lots of noise. Why not only pass the info that 500 tests passed to the AI and only if when something fails add the trace logs?
English
@mattpocockuk If there were a practical way to reduce token usage by 95% with good results u can bet we'd find out from the labs not from a random twitter account
English
@carlosbedia @mattpocockuk Nvm. Some ppl try output as well, but deterministically you only control input
English
@carlosbedia @mattpocockuk I don’t get your point. Input tokens is the only thing you have control over - to change the output tokens you literally have to change the model internals
English
If I had Composer 2.5 in Codex I would be so incredibly happy.
English
Remember when StyleX was going to take over how we write CSS a couple years ago?
English
@rauchg This is what I’ve been noticing as well 😁
Same also goes for setting up your systems and pipelines for speed and rapid iteration. Been pushing for this for a while, now I’m heard because agents need it 😊
English
Agents are motivating so many healthy software habits. Open APIs, documentation (skills), tests (evals), Unix (CLIs), payment & commerce protocols, even wide 𝙰𝚌𝚌𝚎𝚙𝚝 use (markdown/json/html).
The original vision of the WWW coming to life before our eyes.
English
@__roycohen At what scale are you seeing / using this? I feel the same, but in a 500k+ lines codebase I’d like to avoid regular rewrites from scratch
English
@ivanfioravanti IMO the things that defined a software architect in the past. Systems-level thinking. Putting the big pieces into place and having the vision to see the business implications. Knowing which questions to ask and what is load-bearing
English
Code interviews are a thing of the past or not?
You don't ask byte-code or assembler detailed questions. Will high-level programming languages become the same?
What defines a good software engineer nowadays?
🤔
English
@jamonholmgren Did you leave out the loop option on purpose? Where you have -> human prompts the loop -> agent prompts iteratively using the defined loop -> agent ships -> human reviews only critical parts
English
Where do you land in June 2026 on what is the best agentic coding approach? And why?
(I’m trying to keep a finger on the pulse of the industry.)
1. Human prompts, agent codes, ship without significant human review
2. Human prompts, agent codes, human reviews, ship
3. Human codes, agent assists, human reviews, ship
4. No direct agent involvement in coding
0. Other (explain)
English
@felix_dolderer Thanks! It works for web, desktop, and mobile prototypes today
English
Cursor Design is real now.
And we open-sourced it for everyone.
Point. Comment. Mark. Edit. Capture and remix.
Let AI design everything for you. Take control anytime.
This is how a real AI design studio works:👉
Cursor@cursor_ai
With Design Mode, you can now point, draw, or talk to update your UI.
English
Hey @expo can you make a 1-to-1 alternative for eslint-config-expo using oxlint @OxcProject ?
English
@rauchg You guys are shipping crazy amounts of cool new things, 3 of my 4 open browser tabs are just new vercel products to look at/ play around with
English
The hardest part of building an agent is not building the agent. It’s the data.
It’s, ironically, figuring out OAuth, tokens, credentials, scopes… We have AGI sitting there, waiting to be unleashed.
Vercel Connect solves for both security and ease of use. It’s delightful
Vercel@vercel
Vercel Connect makes accessing external data and systems simple and secure. It gives your apps and agents short-lived tokens with precise scopes. vercel.com/blog/introduci…
English
@mattpocockuk Something between confirmation bias and survivor bias
English
I have a deep distrust of almost any 'self-improvement' loop in coding agents
I.e. automatically created memories, CLAUDE.md suggestions applied after every session
Often the suggestions themselves are shit
But even if they're good, the agent often over-indexes on them in a way that's super unhelpful.
It makes the agent impossible to steer. And often because these memories are scoped per-project, each project is unsteerable in its own way.
What's the right name for this? Instruction rot?
English
@bygregorr @ctatedev I think the acronym CORS actually comes from “cursed” and the meaning was only decided after the fact
English
Introducing Generative UI for Claude Code, Codex and Pi
Charts, forms, 3D, anything
Your agent renders real UI for users while it works in a sandbox
Powered by AI SDK's experimental HarnessAgent + json-render
English
@ctatedev I built something like this for onboarding to our codebase. But the thing I built is sooo slow. Even when using Subagents to explore likely next topics. How did you get it so smooth? Is there streaming of structured output and I’m just behind?
English
@Jay_Katana @PromptLLM That is probably the most high-value prompt I use on a regular basis. After the longer conversations, I ask the AI to surface my biases to me so I can make better decisions.
English
@PromptLLM Great for someone who can objectively look at their own life.
However, it’s extremely rare that anyone can look at their own life objectively.
English
