Buenos puntos de @theo. El problema de los últimos modelos es que son demasiado inteligentes/sofisticados para la mayoría de los proyectos.
Las últimas semanas he tenido que reescribir todas las skills para indicar que NO hacer. Menos tests, menos computer use, menos complejidad en el modelo de datos
Theo - t3.gg @theo Nov 17 Why I hate GPT-6.2 Astra During early testing, I found myself so frustrated with this model that I spent a day writing a letter to OpenAI about it. They fixed a small handful of my complaints, but not enough for me to
I came to the conclusion that most SaaS should become an MCP server.
we’re getting the economics of AI software backwards. A lot of AI startups are building their own chat interfaces, agents, and workflows, essentially reselling tokens from the big labs at low to negative gross margins.
What if SaaS companies did the opposite?
Take HR software. Instead of building an AI assistant, become the best MCP server for HR. Expose tools to:
- Benchmark salaries against proprietary market data
- Run compensation reviews and calculate pay equity
- Model salary increases against a department’s budget
- Recommend promotions based on company policies
- Update employee compensation in the HRIS
- Manage approvals, permissions, and audit trails
Let ChatGPT, Claude, or Copilot handle the conversation, reasoning, and orchestration. Let the customer pay for the tokens.
The SaaS company keeps the valuable stuff: proprietary data, integrations, business logic, permissions, compliance, and the system of record.
You still charge a SaaS subscription, but you’re no longer paying to generate every answer. You can ship faster, with fewer engineers, and potentially restore software-like gross margins.
The next generation of SaaS might not have much of a UI at all. Just an incredibly good MCP server.
Sharing this one again, Jensen’s first tweet ever, earlier this year.
“Why open models matter”
Four months later, fired from the company he built from scratch over 30 years.
That a firing of such consequence occurred without a special shareholder vote tells you the real story: it was not about his capabilities in leading Nvidia as CEO, but rather about geopolitics.
And there’s still no evidence the woman was actually a spy, but evidence is an unnecessary precondition.
For my first post, I’m sharing a letter @NVIDIA signed on why open models matter.
AI will transform every industry, power every company, and be built by every country.
Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty.
The world needs both frontier closed models and frontier open models.
images.nvidia.com/pdf/Open-W…
Y si no vuelven?
Es la crítica más frecuente a @exp8fellowship. Pero la mejor manera de quedarse atrás es retener a los mejores lejos de la frontera.
La labor de un pensionado no es volver. Es devolver.
30 HORAS PARA ARREGLAR UNA LANDING QUE NADIE HABÍA MIRADO.
Un partner nos trajo una venta ya cerrada: arreglar una landing en 30 horas. Nadie había mirado el código.
Antes de arrancar, escaneamos la landing con IA desde el navegador. Sin acceso, sin credenciales. Queríamos saber con qué nos íbamos a encontrar antes de comprometer horas.
Apareció algo que cambiaba todo: una pieza clave estaba en un servidor externo, fuera del control del cliente. Eso no se arregla en 30 horas contadas a ciegas.
Vender horas sin mirar ya no cierra. Un pintor viene a ver la pared antes de pasarte un número. Así vamos a seguir cotizando de ahora en más.
BONUS 👇
TIL que puedes crear "Agentes" en Whatsapp que funcionan como los bots de Telegram. Tienen una API key y tus agentes lo pueden usar para mandarte mensajes
Es gratis, a diferencia de los mensajes de Business, y parece que la principal limitación es que solo puedes enviar 12 mensajes por minuto
En @madbuilders_ siempre hemos intentado ser un espacio en el que los que formáis parte de la comunidad pudieseis:
- Descubrir y conectar con otros builders
- Estar al día de las últimas tendencias en tecnología
Por eso hemos construido @bangers_hqbangers.cc 👇
Expanding on why I believe "ETH as store of value (SoV)" makes much more sense to focus on than "ETH as fee-burning token".
Three primary reasons:
1. ETH as fee-burning token never was as strong as it seemed
2. Optimizing for ETH as fee-burning token would demand Ethereum and ETH function in ways that are very hard to make compatible with decentralization and capture resistance.
3. Meanwhile, optimizing for ETH as SoV is aligned with Ethereum focusing on its unique strengths (decentralization and CROPS)
EIP-1559 (including its burn) is fine and good, we should keep them. But we should not feel sad if fee burn is < 10% of breakeven indefinitely; the value of ETH comes from elsewhere.
The core idea: the big thing that people appreciate about Ethereum is decentralization and CROPS. The main pragmatic benefit that these things translate into is stability. SoV is the category where that is not an unneeded extra we merely put effort into for ideological reasons, but rather a pragmatic strength from the perspective of people putting in their assets. And in the era of AI disruption, stability will be highly valued.
If you are in #Madrid , maybe try checking out my Madrid Food Recommendation List :
proximo.to/p/where-i-eat-in-…
(Yes! I designed and developed every tiny detail on the website with Fabien inspired by @kepano . I call it geolocated prose)
After 9 years of blood and sweat building a platform that annually powers:
- $2,000,000,000,000+ in transactions
- 500M+ humans globally
- More API calls than Google does searches
- @jpmorgan@Fidelity@RobinhoodApp and more
we realized our globally scaled infra engine was perfectly constructed to offer ultra low latency, low cost inference
Opening VIP access today.
⚗️ Alchemy Inference is live.
open-weight models. low latency. low costs.
Kimi, DeepSeek and GLM on an OpenAI-compatible API, with zero data retention 🔒
inference.alchemy.com
Re: the Astra modifying Qwen 3.5-0.8B paper today, important clarification.
The story is not "Astra used its LLM X-ray vision to surgically alter the weights of its brethren." It's more like "Astra found a published interpretability technique that can solve this particular kind of failure modes, and implemented it on its own."
It's a cool anecdote, but well within our understanding of LLM capabilities. It's not RSI, and it's definitely not LLM telepathy or whatever people on HN seem to think.
to add some extra clarification here and a small anecdote, looping was discovered as the primary issue after the main result because we originally thought it was just a pure capability gain on math that it had somehow edited in (an even bigger result that isn’t the case).
when Riki and Jayden first presented this result to me, I had a suspicion the large gain in math performance was related to reducing reasoning looping bc I observed a similar effect earlier when training RLMs yielded an odd jump in benchmark scores for a separate paper that was a result of reducing harness-level looping. We discovered Astra indeed used a trick loosely described in arxiv.org/pdf/2406.11717.pdf based on activation steering to reduce this effect
This is also what led to the later ablation about removing looping when the model is told that is the issue but it doesn’t have access to activations.
the tldr; is that the LLM-generated prompt does include hints that Qwen would have this issue, but it started (and evolved without human feedback) as the simple unexpanded prompt without this prior. we will make all of this clearer though (we didn’t even realize until after we got the result that looking for looping was present in the LLM optimized prompt, as we ourselves didn’t even know to look for it initially)
This is just the beginning.
You connect your email to accomplish a specific task. Instead, AI assistants go on and syphon *all* your data
Everything about you, your family, your business
You have no control over it, and most of the time you don't even know it's happening!
This app is insanely farming your entire life when you order this pizza.
I connected my Gmail to order a pizza, and afterward the AI casually admitted it had collected information about my family, closest friends, businesses, travel habits, loyalty programs, favorite restaurants, and even Citi Bike usage.
How fucked up is it that an AI agent tasked with ordering a pizza is digging through your entire personal life?
This is a massive privacy problem with AI agents.
We were able to fine tune an open weight model (Qwen3.5-9B) on proprietary internal data at @coinbase, and it out performed frontier models, at much lower cost and speed, for a specialized task (fraud prevention in this case).
I think we'll see more of this pattern inside companies. The frontier models are amazing generalists. But increasingly you'll see specialist models out perform on more narrow tasks, either working on teams, or by themselves.
We're making it easy to fine tune models internally for our employees. We even have a skill for it in our internal harness that makes it easy to spin these up and start training them:
"/fine-tune-and-serve-llm I want to fine-tune Qwen for payments risk review. help me get started"
I imagine we'll have some backend services that could be replaced entirely by this new architecture.