đ¤ AI builder ¡ Techno-optimist đ Trail runner ¡ Dad of 2 ⥠https://xitter.cf/t.co/O7BBKmgifw đď¸ https://xitter.cf/t.co/zOdNI9j1Jn đˇď¸ https://xitter.cf/t.co/G4w9bcpgZ7 ă https://xitter.cf/t.co/OZApZn4UWN More at: https://xitter.cf/t.co/Atjn14ZNnq
Catalonia
Joined May 2008
- Tweets6.8K
- Following1.2K
- Followers919
- Likes2.7K
Pinned Tweet
Twenty years in, and the engineering only gets more interesting.
I still write production code. That is the job, not a side habit I mention in interviews. I use AI to help me code, review, and ensure all my code is secure and documented; I'm not afraid to say that.
Most weeks, I design AI solutions that have to run on a companyâs real knowledge: what they already know, what they are allowed to use, and what the model isn't allowed to invent. Prototype-to-production is the interesting part. I also hire, and I still take the call when a system is on fire.
What that looks like in practice:
At Keeeb I stood up a remote engineering team from zero, 15+ people across backend, AI, and frontend, then took a monolith apart into 12+ Kubernetes services and a public API SDK. The interesting part was not the org chart. It was scaling a product without turning it into a science project.
At Medullar I built the first proof of concept myself, then took a privacy-first, knowledge-powered AI platform the rest of the way: how the knowledge gets in, how the models use it with RAG, agents on top, the production path. I will happily live in that problem until the design is boring in the good way.
I also run a small studio, Luodaint, and ship my own products (Corvidly, Feedjolt, InfluCodes). I like 0-to-1. I like a team. I like AI when it changes the system, not when it is a badge on a changelog.
I am a poor fit if you want a CTO who only manages, a culture of calendar Tetris sold as speed, or a mandate to âdo AIâ with no opinion about the underlying product. Startup, hands-on, real problem. That is the seat.
If you are hiring a founding engineer, Head of Eng, or a fractional CTO who will still open the repo, send me a note. Coffee is fine too. I am in Terrassa; I work remotely, and I have done this from Barcelona, Dublin, and Boston.
Liquid AI shipped LFM3: six open-weight models on one 1.7B backbone, 12.7x better on agentic tasks than their last 2.6B model. Architecture beat scale. Fine-tunes on a single consumer GPU - the lane solo founders live in.
SundayQuill's best search day ever was 19 Oct: 26 impressions, average position about 1.7. The whole next week: 2 impressions.
That is what a search footprint of a handful of pages looks like. Demand shows up when the brand surfaces, then vanishes because there is almost nothing to find.
Launch content is how the spike becomes permanent. Product Hunt in 27 days.
Free to read in the meantime.
ÂżTe gusta leer? ÂżQuieres tener un ratito de entretenimiento y desconectar?
Mira sundayquill.live una web con mĂşltiples libros que se van escribiendo cada semana. Como usuario, puedes sugerir cĂłmo te gustarĂa que se desarrollara la historia.
Pruebalo, es gratuito.
Anthropic runs its own drug discovery: rare childhood muscular dystrophy, Claude on much of the science and ops, plus 6 pharma partners. Labs becoming operators is the tell. anthropic.com/research/earlyâŚ
Productboard shipped ARR columns and custom impact scores this month.
That closes the "we rank by ARR" era of feedback tools. When the category leader makes a differentiator a standard feature, it stops being a differentiator. The fight moves to the pricing model: per-seat feedback tools tax the teams closest to the customers, so volume gets rationed.
Feedjolt: flat rate, unlimited contributors, collects Slack, email and tickets with semantic dedupe.
Startup $9/mo, by application. Growth $15/mo, billed annually. Scale $39/mo, billed annually.
Genie 4: DeepMind's world model redraws a simulated scene live when you ask for rain or a new character. Audio too. SundayQuill is the words-only take on that. deepmind.google/blog/genie-4âŚ
The tokenizer tax: even among the 20 most-used languages online, the most expensive costs 1.5 to 5x more tokens than the cheapest. When the study expanded to 200 languages, the gap hit 17x in one model.
That lands as higher cost, higher latency and less usable context for non-English users, decided at the tokenizer, not at your product layer. Script matters more than speaker count, so a language with millions of users can still get the worst rates. If you price per token, you are pricing your market by script.
Cloudflare disclosed that 155 responses on Workers AI contained other customers' data between August 27 and September 9. Not a prompt trick: the KV cache pool was smaller than the advertised 1M-token context, and an SGLang bug let rejected requests reach decode anyway. Cloudflare states it found no evidence of exploitation.
An advertised context window is a capacity promise. If the cache behind it cannot hold that much, the overflow goes somewhere, and somewhere is someone else's session.
Today I read this, and I thought it made a lot of sense:
âYou donât need to code.
But you do need to know how to code.â
Itâs like addition. We donât need to add numbers ourselves; calculators do that. But if you donât know how to add, youâre in trouble.
Wikipedia's AI-cleanup began in 2023, spotting clichĂŠs like 'stands as a testament.' AI text hides better now. Honesty doesn't: SundayQuill is AI-written and says so. spectrum.ieee.org/wikipedia-âŚ
new iPad mini starting at approx. ⏠750. in Spain
the joke is that if you want it to fold, thatâll be âŹ2300 approx
Apple has officially announced the all-new iPad mini! đ¨
⢠Larger 8.5-inch OLED display
⢠A20 Pro chip
⢠Up to 1,200 nits brightness
⢠Optional nano-texture display
⢠Thinner and lighter design
⢠IP67 water and dust resistance
⢠New Acoustic Display speakers
⢠12MP landscape Center Stage camera
⢠Wi-Fi 7 and Bluetooth 6
⢠New C2 cellular modem
⢠Fast charging support
⢠Apple Intelligence and Siri AI
Starts at $699
Pre-orders: October 28
Available: November 6
Brought my car in for its annual checkup. Meanwhile, Iâm having coffee and doing some remote work on a nice terrace (perfect weather in Barcelona today).
OpenAI's model claims it refuted the Exponential Time Hypothesis â 1999 conjecture behind hundreds of hardness results. Found Oct 16, Lean-checked a day later. Its proposers lead the review. news.ycombinator.com/item?idâŚ
