AI Safety Researcher. Blogs at https://xitter.cf/t.co/lW4RVHkaHe

Lisbon, Portugal
Joined April 2007
Alcides Fonseca retweeted
A ideia de que grandes proprietários não devem participar no debate sobre habitação na Assembleia da República é profundamente antidemocrática. Apesar de todas as nossas diferenças, acho que o PCP tem todo o direito a apresentar as suas propostas.
88
317
37
2,639
45,002
Alcides Fonseca retweeted
Ao não ser preciso entrada, mesmo com o aumento do preço, facilitou a compra da casa por parte de jovens imigrantes e nacionais. Mas para todos os outros ficou estupidamente mais caro. Se o Montenegro não renovar os apoios que deu, as casas não se aguentam nos preços atuais.
Se ter casa já estava caro com o PS, seguramente ninguém precisava do desastre que se seguiu. Uma aceleração vertiginosa de 22% face à tendência anterior. Por 100m2, um jovem paga hoje mais €42.240 graças a Montenegro, com garantia dos nossos impostos. Vá, não sejam egoísticos😉
7
1
14
953
IMHO, all three major harnesses (Codex, Claude and Cursor) need a Snow Leopard release. Two months dedicated to making the current features working spotless. No slow behaviors, no disappearing messages, no random isolation between threads/projects, clear observability.
1
6
352
Same in Portugal. Except that it’s 11:22 of the same day… two years later.
At 11:10, I sent an email to a local government entity. At 11:22, I received a response which addressed all questions. This is something I like about Switzerland 🇨🇭
2
9
573
Alcides Fonseca retweeted
How long it took me to find each SWE job: 2014: 1 job applied, 0 interviews, 1 day (nepotism) 2015: 1 job applied, 2 interviews, 1 week 2016: 1 job applied, 4 interviews, 1 week 2016: 1 job applied, 0 interviews, 1 day (friend is a CEO) 2018: 1 job applied, 2 interviews, 1 week 2019: 1 job applied, 5 interviews, 1 week 2021: 1 job applied, 0 interviews, 1 day (invited) 2022: 7 jobs applied, 16 interviews, 3 months 2026: 167 jobs applied, 54 interviews, 1 year 4 months
142
137
39
4,335
376,172
So apparently it exists now: rstation.ai but they are not yet orchestrated as in Sakana's Agentic Scientist (pub.sakana.ai/agentic-scient…). Interesting times.
Remember SETI or Fold@Home? We need OpenSource@Home, where we use the tokens that would expire in everyone's subscriptions to contribute to the public good.
1
4
329
Alcides Fonseca retweeted
We are hiring! I am looking to hire a PhD student to work on topics around AI evaluation from an HCI and technical standpoint. If you are interested, please apply through the Mila portal and put my name down as a potential advisor:
Mila's annual supervision request process is now open to receive MSc and PhD applications for Fall 2027 admission! For more information, visit mila.quebec/en/prospective-s…
13
53
2,492
New post on what I learned writing a proxy MCP that allows you to write multi-call policies to block data exfiltration and other harmful sequences of (individually harmless) calls. wiki.alcidesfonseca.com/blog…
1
8
444
Why should we hire you when Claude Code is better at writing requirements and architecting solutions?
You're in a tech interview. The interviewer asks: "Why should we hire you when Claude Code can already write code?" What's your answer?
2
2
615
And if you’re in Java world, checkout LiquidJava.
My favorite talk from Scala Days 2026 was about baking refinement types right into the compiler. Unlike libraries like Iron or Refined, it adds a built-in solver that verifies logic and math across multiple variables at compile time with zero runtime cost. Moving from simple checks to full verification is how Scala's type system should evolve to constrain LLMs - AI code won't compile unless the math holds up.
1
14
656
Alcides Fonseca retweeted
Replying to @simonw
Everybody saying "this is just a source of revenue for Anthropic!!!" isn't thinking about the business lost from their last-gen models being near-free
2
1
14
1,540
Dumbass comment. Linus built it in his garage. Anthropic converted a metro into a Shinkansen mid-trip.
Anthropic, with the help of Claude, spent five months building a replacement for Git, while Linus Torvalds built Git single-handedly in just 10 days.
7
521
Alcides Fonseca retweeted
We're exploring different methods of generating documentation for test suites and need your help to evaluate them! If you have 5 minutes, please fill out the following survey: onboarding-survey-7ko.pages.… (also, please share this with your own network)
1
1
203
Dumb question: OpenAI in Azure doesn't let you use structured_output with parallel jobs. Any idea why?
1
304
When you find out it's lunch time.
7
279
Hire me instead. None of those Google bad habits!
Any startup that hires enough ex-Googlers ends up with an ad hoc, informally-specified, bug-ridden, slow implementation of half of Google infra
1
8
849
I wake up to this? I am very confused.
1
7
442
Models can hide functionality in its weights until a password is entered: wiki.alcidesfonseca.com/blog…
3
349
Alcides Fonseca retweeted
Pus o Jev da @typesafeai contra o Clef e o Clef-flash da @Cloudflare em 300 jogos de xadrez, com o Stockfish (o motor de xadrez de referência) a avaliar cada jogada. O Clef ganhou ao Jev por 57 a 43, e o Clef-flash empatou com o Clef a um terço do preço. Para quem constrói com estes modelos, o número mais útil é a confiança: acima de 0,8, o Clef jogou a jogada do Stockfish 98% das vezes, em 112 jogadas. O Jev chegou a esse nível 4 vezes em 100 jogos. Entre 76 e 83 de cada 100 jogos acabaram empatados, quase todos por repetição (a mesma posição três vezes). Estes modelos leem a posição uma vez e escolhem sem planear. Perante a mesma posição dão sempre a mesma resposta, e as peças andam para trás e para a frente até ao empate.
1
2
6
359
New paper about how we converted English documentation into Liquid Types using a DFA-guided agentic workflow: wiki.alcidesfonseca.com/rese…
3
7
1
35
1,627