Mariano Falcón

Grateful to be alive and listen to everything @mjkeenan sings.

@andavip Lo traté mal a Claudito y resulta que sólo en una de las queries de Metabase había un filtro por hora además de la fecha.

La nueva forma de programar es esta

No puede ser legal publicar esto...

@gptcrosa @bryan_johnson Lo que deben ser esas erecciones a la noche.

Que locura @bryan_johnson 30mg de dmt es una barbaridad 😭

@aperfectcircle Argentina guys, Argentina.

Tickets on-sale now: 4 Dec – Adelaide – The Drive 6 Dec – Melbourne – Rod Laver Arena 8 Dec – Brisbane – Riverstage 11 Dec – Sydney – TikTok Entertainment Centre 13 Dec – Auckland – Spark Arena Puscifer join us on all dates. 📷: Travis Shinn

¿Qué tendrá que ver el muchacho con una pelota en la descripción de un crédito hipotecario?

two weeks ago, my sink was clogged in the past, I would have called a Plumber. but this time, I used gemini live. It guide me, told me what to buy/what I needed and what could be happening. To make sure, I cross checked it with opus - 15 minutes approx result: sink fixed. now I understand how it works and I can fix it next time if it happens again. let that sink in…

Vector databases are dead!!

The human LLM enthusiast urge to declare that X is dead: - Open AI has launched Y! 384934893 startups are now dead. - RAG is dead! - MCP is dead!

@TomasPiaggio @DevvMandal @markov__ai

@DevvMandal @markov__ai watching 10,000 hours of salesforce recordings sounds like a very specific circle of hell.

Today, we're launching the world's largest open-source dataset of computer-use recordings. 10,000+ hours across Salesforce, Blender, Photoshop and more, to automate the next level of white-collar work. Link in the comments :) @markov__ai

@ElEconomista_ Si la mujer compraba un pasaje por una aerolínea comercial el estado no recibía dinero por los impuestos de la transacción?

🟥''Yo vengo una semana a deslomarme y quería que mi esposa me acompañe'': Adorni defendió el viaje de su esposa a Nueva York en el vuelo presidencial y aseguró que ''no le saca ni un peso al Estado''

Was using gpt-5.2 model in a conversational AI chatbot, then I switched to gpt-5.4 and the edge cases that were failing got fixed without any code changes.

@impronunciable @evilrabbit_ Me hiciste acordar del edificio donde se hizo la Mozfest 2016 y lo tuve que googlear porque estaba muy copado, Ravensbourne College.

@evilrabbit_ Si podes date una vuelta por el Barbican Centre

London, your vibe is unique.

Falconius predijo OpenClaw

@JnBrymn I think that to remove the huge code review bottleneck we have to focus on the design of better verification artifacts: tests, evals, security analysis, stress tests...

Building software with AI will be solved in less than a year – if it isn't already basically solved now! Rational: - A dwindling minority of naysayers (poor souls at large/old/slow-adopting companies) are pretending that AI is still bad. - A growing majority is rapidly adopting AI in their daily dev experience. - A large percentage of them is actively working on making their AI setup better - and the good ideas are really sticking around! Top of mind things left to resolve: - Most work is still single threaded. Swarms of AI assistants would work faster, but coordination is still challenging. - Transparency of work – It's still a little too easy for AI code to "get away" from the developer where they no longer have a firm understanding of the design so that they can make intelligent requests. - Having the AI challenge assumptions – AI too often makes assumptions about what you want when sometimes it should raise a flag and just ask about critical design decisions you've omitted. - Large, pre-existing codebases are still more challenging than green-field projects. - Guardrails in implementation – making sure some invariant isn't ignored and the corresponding tests deleted! - No one has figured out code review yet – there is now too much code to review line-by-line. So should we start reviewing intent and double checking guardrails? What do you think! I WANT TO KNOW!

Come on Opus 4.6, you can do better...

I customized CC spinner verbs to use "It's always sunny in Philadelphia" words, and now CC says it's doing Charlie Work. "Charlie Work is the janitorial work that needs to be done at Paddy's Pub and that most of The Gang refuses to do. It is typically done by Charlie Kelly."

@honnibal Yes, I ask CC to never modify a test unless specified. I always try to decouple coding logic and tests iterations.

Does anyone else find Claude Code incredibly sneaky as soon as you get a test failure that might actually matter? I have to fight very hard to make it investigate instead of just redefine the test and say it's supposed to be like that

I've written about this in the following post: aunhumano.com/index.php/2026…

Verifying the AI coding agents work is a crucial stage in this new dynamic and we have to invest time to get it right. For traditional deterministic software you have to work hard on tests but if you're working on non deterministic software you've to also work hard on evaluations

The code written by the AI coding agents it's getting better and better. They can write a lot of working code in a short amount of time. Manually reviewing all the code it's now the bottleneck to really get this new workflow to scale.