Juan retweetledi
Juan
3K posts
Juan
@ditorodev
building better and fairer interviews at https://t.co/Eb1NuUwExM
Barcelona Katılım Eylül 2020
2.5K Takip Edilen365 Takipçiler
It's 2026 and your CEO just sent you a 2,400 line pull request.
You get a cup of coffee and sit down to review it.
It's a disaster. A dozen unrelated refactors. Unused methods with names like `convertFromBase10` and `normalizeBeforeSerialization`.
You catch a few hardcoded API keys, but that's ok. It's part of the dance. They didn't consider that someone might look at this diff. Here's a comment buddy.
They respond in an hour (after Copilot, qodo, CodeRabbit and Greptile finish their reviews) saying we shouldn't worry about "implementation details" anymore, those are relics of the past. Hey let's jump into a room and figure it out. We can't just agree to disagree, this is probably my last job in tech and I can't watch this fucker burn the place to the ground.
The PR merges and goes to prod. You feel a shared sense of apathy and dread with Hannah the intern (she has to review his AI generated social media posts ever since Grok got too imaginative).
That night you go to sleep and have nightmares of that code. You can still see the shapes of it on the backs of your eyelids.
You go to work the next day ready to quit. You no longer understand the system. There is no foundation. Time to use those savings and an SBA loan to buy a liquor store and never login to GitHub again.
staysaasy@staysaasy
It’s 2018 and your coworker just sent you a 400 line pull request. You get a cup of coffee and sit down to review it. It’s beautiful. Elegant micro-refactors. Crispy method names. You catch a few things, but that’s ok. It’s part of the dance. They didn’t consider extensibility on part of their API. Here’s a comment buddy. They respond in an hour saying they think we should do one piece differently than your comment. Hey let’s jump into a room and figure it out. We can’t just agree to disagree, this code is too important. The PR merges and goes to prod. You feel a shared sense of ownership and accomplishment. That night you go to sleep and dream of that code. You can still see the shapes of it on the backs of your eyelids, your IDE syntax highlighting sparking neurons in your reptile brain. You go to work the next day ready to go. You understand the system. N is your foundation. Time to build n+1.
English
@jeremyphoward They could 100% build a better way to get all this training data on an harness that’s not shait
English
@jeremyphoward All this is is “help me improve my model thru Claude code and I’ll subsidize you or get lost”
To pay 200$ and get 200$ I had rather just not subscribe
English
This is misleading.
This policy redefines the term "interactive" to mean "using an Anthropic front-end".
If you use `claude -p` or Agent SDK to do something interactively, it now uses credits, not your subscription limits. So the "interactive use" heading saying "unchanged" subscriptions is not accurate.
Lydia Hallie ✨@lydiahallie
To add some clarity: you don't pay extra. It's the same subscription, same price per month. What's new our sub now covers two separate pools: · Interactive → sub limits, unchanged · Programmatic → new $20–$200 included(!!) credit, metered at API rates
English
@jarredsumner no fucking way that in 6 days the rewrite is done, this shit crazyyyy
English
Bun v1.3.14 releases tomorrow.
If we do merge the Rust rewrite, this would be the last version in Zig
English
Juan retweetledi
Juan retweetledi
Today is a hard day. I shared this note with the @linear team today: We’ve made the difficult decision to increase our workforce. This is not a cost-cutting exercise or a reflection of anyone’s performance. We’re simply reimagining every role for the agentic AI era. We’re hiring. We’re sorry about that.
English
Juan retweetledi
Pasado mañana Jueves en BCN (7 Mayo) de 10 a 12: ME PONGO PINGANILLO: >>> MUST: Ven sólo si eres reclutador o co-founder reclutando (sino vas a tirar tu tiempo a la basura) 1/3 >>
Español
@kacemmmmmm @GeoffreyHuntley @glcst The most productive I have been with Python has been with the Python debugger open at all times but it means I’m only able to work one feature at a time
Things like effect you can design an ecosystem that will keep your agent ahold and around your system quickly
English
@kacemmmmmm @GeoffreyHuntley @glcst I’m using all of these and still the experience is pretty bad. The typing system is super incomplete and inflexible and pydantic alone sadly cannot do the job
English
good point; won’t disagree there re: people continuing what they know/do without rethinking things and network effects.
idk if library argument matters anymore. pre ai i wouldn’t even consider ocaml/haskell because ecosystem effects but these no longer matter. actually picking a lang due to open source ecosystem that one can consume is now an anti pattern tbh.
re: “software”
sad but true
English
@kacemmmmmm @GeoffreyHuntley @glcst These two:
- Weak type system
- Slow feedback loop with the agent
- Lack of correctness frameworks, with how wild agents are im missing something that forces correctness proactively like Effect on typescript
English
@GeoffreyHuntley @glcst yeah im already seeing those needs
(im here planning 1 day of work to iterate on our own implementation of github.com/getsentry/ward… to see if I can tame them a bit)
English
@ditorodev @glcst but indeed, unless you need python (ML topics) then ditch it. general recommendation i give to folks these days. it won’t get you far. you’ll have to do more engineering to keep agents on the rails.
English
Would you buy Zed Pro if we added open source models from our friends at @baseten
Amario@amariokoro
@zeddotdev why not add open source models from american inference companies to the subscription you provide, that way the 10$ would actually be worth it espcially for students
English
@NielsRogge @MistralAI wait what mistral medium 3.5 is better than m2.5?? wtf
English
People may bash on @MistralAI...
...but it's also the only non-Chinese model in the top 25 (!) of open models on SWE-Bench Verified
English
Qwen3.6 27B landed yesterday, so I ran it on 4 setups side-by-side to see how they stack up:
🔴 RTX 4090 — 45.59 tok/s, TTFT 525ms
🟢 RTX 5090 — 51.83 tok/s, TTFT 752ms
⚫️ M2 Ultra — 22.30 tok/s, TTFT 216ms
🟣 DGX Spark — 11.08 tok/s, TTFT 319ms
This is a standard test: no tuning, just the out-of-the-box experience.
For the NVIDIA cards I used llama.cpp with Unsloth's UD-Q4_K_XL quant. For the M2 Ultra I used MLX with Unsloth's UD-MLX-4bit quant, since MLX is the native path on Apple Silicon.
Please consider this as the baseline, you can definitely squeeze more out of every one of these with fine-tuned settings.
English
Postgres has three ways to isolate tenants:
- Logical databases
- Per-tenant schemas
- Tenant ID in a shared schema
Counterintuitively, the last is the best way to scale. Read about why in our latest article.
English
Juan retweetledi
Extraordinary scenes on my TL
Guys there’s more than one way to optimize AI on a task. If you’re working on harnesses try to slowly add all these in your bag.
The classic way is to update the weights (RL)… The modern way is to optimize prompts/context (Dspy optimizers/GEPA)… and the hypermodern way is to self evolve the codebase itself (auto research/alpha-evolve/darwin-godel variants)
All of them need an eval dataset of prompts/task scenarios, a rubric of success, and an initial forward pass (harness+model) to learn. They just update different things to get your system to better evals.
There’s nuance to each. There’s a time and place for all of them.
English
Gotta start wrking on getting familiar with this. Pretty sure this is also how you get better results with cheaper and faster models
Sam Hogan 🇺🇸@samhogan
RLMs pretty much solved context btw You can shove tens of millions of tokens into a good RLM harness and it just works. I’m spending all my free time here.
English