CTO & Co-Founder @LayerXlab. Building @taikainetwork, book author and engineer.
Oporto, Portugal
Joined May 2009
- Tweets1.8K
- Following1.8K
- Followers815
- Likes3.4K
Pinned Tweet
After months of putting it off, my personal website is finally live: hvasc.dev/
The code is open source, so feel free to fork it and make it your own.
𝗪𝗵𝗮𝘁 𝗱𝗼𝗲𝘀 𝘁𝗵𝗲 𝘄𝗿𝗼𝗻𝗴 𝗺𝗼𝗱𝗲𝗹 𝗳𝗼𝗿 𝘁𝗵𝗲 𝗷𝗼𝗯 𝗰𝗼𝘀𝘁 𝘆𝗼𝘂?
Yesterday I spoke at the API meetup at Porto about choosing between open, closed and open-weight models.
Here's the short version (slides attached):
✅ Scores cluster. Costs don't. On the latest Artificial Analysis Intelligence Index, six frontier models sit between 52 and 58 points, while their cost per task ranges from $0.72 to $7.63.
✅ Decision models are the quiet win.
✅ Route by the cost of being wrong.
Your model choice is no longer a single pick. 𝗜𝘁'𝘀 𝗮 𝗿𝗼𝘂𝘁𝗶𝗻𝗴 𝗽𝗼𝗹𝗶𝗰𝘆.
Helder Vasconcelos ⛩️ retweeted
Two AI launches in three weeks. Is either one AGI?
Depends who you ask. Our CTO @heldervasc explains why, and proposes a simpler test: stop saying "good, for an AI."
By that test, maths and code are already there, science is close, and the physical world still has a way to go.
Find more. 👇
Helder Vasconcelos ⛩️ 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.
Now the only limit is creativity.
I left the music industry years ago with a drawer full of instrument and effect ideas I never had the means and time to build.
Time to open that drawer 🎉 @Re_Axis 🤫
Artists and producers can now create their own custom instruments and effects, then export them as VST3/AU plugins to use directly in their DAW.
We're making it easy to bring these capabilities into professional setups, so producers can easily build tools tailored to their creative process.
Read more: blog.google/innovation-and-a…
Helder Vasconcelos ⛩️ retweeted
Aaaaand we’re back for yet another internal hackathon!
There’s no better excuse to gather the team than a one-day competition to see who can build the most useful project.
We’ve done dozens of internal hackathons over the years, and it’s still as exciting as the first one.
Time to build!
𝗪𝗵𝗮𝘁 𝗶𝘀 𝗔𝗚𝗜 𝗳𝗼𝗿 𝘆𝗼𝘂 ? 💭
Everyone is selling AGI. Almost nobody agrees on what it is.
OpenAI defines it by jobs. DeepMind uses capability levels. The Microsoft and OpenAI contract once measured it in profit. Each definition fits the incentives of whoever wrote it.
The capability is real, though. In April, GPT-5.5 solved about a third of FrontierMath Tier 4, which is research-level mathematics. GPT-6 Astra now solves around 95%. The same models still struggle with spatial reasoning.
So I use a simpler test: when a result no longer needs the qualifier "...for an AI," that field has crossed over. Math has.
Code is close. Robotics isn't there yet.
𝗔𝗚𝗜 𝗶𝘀𝗻'𝘁 𝗮 𝘀𝗶𝗻𝗴𝗹𝗲 𝗺𝗼𝗺𝗲𝗻𝘁. 𝗜𝘁'𝘀 𝗮 𝗱𝗶𝗮𝗹, 𝘁𝘂𝗿𝗻𝗶𝗻𝗴 𝗮𝘁 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝘁 𝘀𝗽𝗲𝗲𝗱𝘀 𝗶𝗻 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝘁 𝗳𝗶𝗲𝗹𝗱𝘀.
Check out my full post on @LayerXlab blog 👇
🇬🇧 English Version: layerx.xyz/blog/what-is-agi
🇵🇹 Versão em Português: layerx.xyz/pt/blog/what-is-a…
Helder Vasconcelos ⛩️ retweeted
The SpaceXAI Digital Transformation Lab is happening on TAIKAI!
Turn a real business challenge into an AI solution in 48 hours, on 26-27 October, in Fórum Braga.
Powered by @IAPMEI and supported by @StartupPortugal, the applications open until 16 October. Come alone or bring your teammates!
Turn a real business challenge into an AI solution in 48h.
Applications are open for the SpaceXAI Digital Transformation Lab, a hackathon for SMEs.
📅 Apply by 16 Oct
📍 26–27 Oct | Fórum Braga
🇬🇧 In English
Apply 👇
iapmei.pt/pt/paginas/laborat…
Helder Vasconcelos ⛩️ retweeted
Does your Claude model really need Claude Code…? 🤔
We evaluate 7 models on Claude Code, Codex, and Pi. Three surprising findings emerge:
1️⃣Harness choice has little effect on task success rate, but can significantly affect the cost
2️⃣A simple harness can be competitive
3️⃣The native harness isn’t always the best.
Millions of people are using coding agents, but the impact of harness choice remains unclear.
(1/n) More details in the thread. 🧵
𝗪𝗛𝗔𝗧 𝗜𝗙 𝗧𝗛𝗘 𝗔𝗜 𝗝𝗢𝗕 𝗗𝗢𝗢𝗠𝗘𝗥𝗦 𝗔𝗥𝗘 𝗪𝗥𝗢𝗡𝗚?
In January at Davos, we were told AI could do "most, maybe all" of what software engineers do within 6 to 12 months.
My feed filled up with "developers are cooked." 😱
Eight months later, here's what the data shows:
✅ US software developer job postings are up ~15% since February 2025, while overall US postings fell 7% (Indeed Hiring Lab)
✅ A Federal Reserve paper found more people held software developer jobs in 2026 than in 2025
✅ In the recent rebound, the most AI-exposed occupations recovered the most
✅ BLS still projects 15% growth for software developer jobs through 2034
The jobpocalypse crowd got this badly wrong. So far, AI looks far more like a tool that expands what companies build than one that replaces the people building it.
To be fair, it's not all good news. Developer postings are still ~27% below pre-pandemic levels, and 71% of the recent growth went to senior roles. Juniors are having a genuinely hard time. But that slump started with the 2022 overhiring bust, before ChatGPT even existed.
Bold predictions move headlines and markets. Data moves slower, and right now it's telling a very different story.
Stop panicking. Keep building. AI is not going to replace you is a tool to improve your productivity.
Helder Vasconcelos ⛩️ retweeted
This September, we're building at @dhworkspaces Picoas.
A space made for exactly this kind of day: room to spread out, focus, and work alongside 100 other builders shipping real AI projects.
September 25, Lisbon, during @lisbonaiweek.
Helder Vasconcelos ⛩️ retweeted
Choosing an AI model is no longer just about picking the one at the top of a benchmark.
@heldervasc breaks down how to actually choose the right AI model, at The Collider on October 15.
Save your spot. 👇
luma.com/qlvrf8sl
Helder Vasconcelos ⛩️ retweeted
There are two AGI scoreboards. One is capability, where the frontier still leads. The other is cost per task, where open weights now win by up to 50x.
Picking a model is the wrong move. The fix is a routing policy: cheap-to-be-wrong calls go to self-hosted open weights, expensive-to-be-wrong calls go to the frontier. @heldervasc breaks down why.
It is August and I have lost count of the frontier releases this year.
I sat down to pick the one we should standardise on. Never got an answer. Got a better question.
Model choice is a routing policy now, and it is yours to build
hvasc.dev/articles/agi-two-s…
Helder Vasconcelos ⛩️ retweeted
What does a day of building with AI actually look like? We put it on camera.
From first-timers to seasoned devs, we asked people at AI Hacker House: Solstice Edition what brought them to Porto, and @hypermarioal shared why LayerX put this together in the first place.
Hosted at @dhworkspaces Península, with prizes powered by @YLDio and @supabase, and Sumol Compal, Pleno, Vital Shot, Padaria Formosa, Braza and RHEA Therapies keeping everyone fed, hydrated and relaxed. Thanks also to @lisbonaiweek, @TalentProtocol and Innovation X Hub for helping spread the word.
Watch the recap below. ⬇️
