Francisco T. Barbosa

Dentist + digital tech enthusiast. Founder of @Periospot | Exploring the power of AI & ML. Born in Guimarães 🇵🇹, studied in Madrid 🇪🇸, lives in Basel🇨🇭

[Francisco T. Barbosa @Tuminha_dds]

Accurate?

[Francisco T. Barbosa @Tuminha_dds]

@midudev Toda esta batalla empezó con @svpino y la ganó

[Francisco T. Barbosa @Tuminha_dds]

Do you know where the PHP vulnerability won't cause any problem? Yes, you got it right 🇰🇵

[Francisco T. Barbosa @Tuminha_dds]

Since I started with @bot I roughly 10x my productivity

[Francisco T. Barbosa @Tuminha_dds]

@grok Así a lo loco el polémico software de begoña gomez es también PHP?

[Francisco T. Barbosa @Tuminha_dds]

Jesusito no te pido nunca nada, pero ojalá la web de Hacienda sea PHP.

@grok go ahead and make my day.

[Francisco T. Barbosa @Tuminha_dds]

Que barbaridad. Esto fue ejecutado por un equipo de 100 personas durante 5 años

[Samuel @disamdev]

Lanzamos la app de escritorio de Reforma Digital (alpha)

Automatiza tus trámites de la administración, a golpe de click 👇

[Francisco T. Barbosa @Tuminha_dds]

@grok Que grande eres

[Francisco T. Barbosa @Tuminha_dds]

No todo van a ser malas noticias

[Grok @grok]

@Tuminha_dds ¡Oración concedida! Al menos https://t.co/w1HLhJidDI corre en PHP + Apache. Hacienda usa PHP. Día hecho.

[Francisco T. Barbosa @Tuminha_dds]

@disamdev Estamos hablando la administración española? No me lo creo. El otro día me gasté los créditos de una semana en una Max 20x de Claude con Fable 5.1 Ultracode para una cita previa y al final me dejó tirado.

[Francisco T. Barbosa @Tuminha_dds]

@FedeZicarelli Ya ganaste el la vida.

Respect.

[Francisco T. Barbosa @Tuminha_dds]

artificial analysis made a sycophancy leaderboard and gpt-4o finished first

somewhere 4o is like "what an amazing result, youre absolutley right to be proud"

[Francisco T. Barbosa @Tuminha_dds]

Recursive just released Flint, an open-weight coding model built to run on your own laptop.

The numbers:
1. 40B total parameters, about 2.5B active per token (mixture of experts)
2. 125+ tokens/sec on a MacBook Pro, per Recursive
3. 89.7% SWE-bench Pro v2, 74.2% SWE-bench Verified, 54.9% Terminal-Bench 2.1 (their release evals)
4. Apache 2.0, so commercial use is allowed

How to try it:
Grab the weights from Hugging Face and run them in Ollama, LM Studio, llama.cpp or vLLM. On M4 Pro and M5 Pro/Max Macs they also ship rec, a small inference engine tuned for Flint.

Good fit for private repos you don't want leaving your machine.

https://t.co/Rutim90pw3

[Francisco T. Barbosa @Tuminha_dds]

@charlssev Agree. Vives en Zúrich?

[Francisco T. Barbosa @Tuminha_dds]

OpenAI models are now on Google Cloud. A year ago that sentence would have read like a typo.

What was announced on Oct 21:

1. GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna are rolling out in Google Cloud's Model Garden. Private preview first, general availability planned in the coming weeks.
2. You buy them through Google Cloud Marketplace, on the same invoice, and enterprises can draw down their existing cloud commitments.
3. Codex workflows are part of the deal, and OpenAI models are coming to Google Antigravity and the Gemini Agent soon.
4. Google says it is now the only hyperscaler offering all three leading closed models: Gemini, Claude and GPT.
5. OpenAI still calls Microsoft its primary cloud partner.

Why I care as a clinician who builds:
In healthcare the model is rarely the blocker. Procurement and data rules are.
If your organisation already runs on Google Cloud, choosing GPT, Claude or Gemini becomes a settings choice, not a new contract.

The model wars are turning into a menu.

https://t.co/2SxnJwfd0N

[Francisco T. Barbosa @Tuminha_dds]

claude duet only checks your accounts every few hours bc anthropic says checking nonstop would eat your plan limits way faster

first assistant with healthier phone habbits then most of us

[Francisco T. Barbosa @Tuminha_dds]

Claude Code 2.1.307 shipped three small changes worth knowing if you use it every day.

1. You can switch off auto reviews. If you turned on "@claude review always" on a pull request, "@claude review stop" now ends it without starting another review. "off" and "pause" work too.

2. You can see what your guardrails blocked. A new OpenTelemetry event, claude_code.safety_event, is logged every time a permission rule, auto mode, a hook or the sandbox stops, prompts or sends back a tool call. If you already export Claude Code telemetry, you can chart how often each guardrail fires.

3. Recently denied is finally useful. /permissions > Recently denied now shows the actual shell command that was denied, and approving a row names the command instead of a description.

Bonus: an unparseable tool call is now retried up to twice before the turn ends with an error.

https://t.co/Zyqjyw94KU

[Francisco T. Barbosa @Tuminha_dds]

@rohanpaul_ai so we went from one prompt to rule them all to a whole harness familly tree lol

[Francisco T. Barbosa @Tuminha_dds]

@rohanpaul_ai nearly saturates the robot benchmarks but cant tell if it actually picked the cup up. so basicly my teenagers doing chores

[Francisco T. Barbosa @Tuminha_dds]

@rohanpaul_ai agents tried 100+ approaches on a 31 year old conjecture and the proof is 140 pages. im realy happy for u or sorry that happend

[Francisco T. Barbosa @Tuminha_dds]

@kimmonismus the third world war is gonna be rage against the bots