Founder, ECHO Consulting + Restrained Depth AI. Licensed clinician helping implement AI workflows without losing human judgment. 225M+ views building in public.
California, USA
Joined March 2026
- Tweets2.9K
- Following2K
- Followers4.2K
- Likes10.6K
You watched us build them. 200M+ views on X. Now open them: a resume from a sticky note, a live voice interviewer, college applications decoded, a live assistant for your business. Free with an email in 4 languages.
Preferred Networks cuenta que, al desarrollar PLaMo 3 Translate, exploró un método que mejoraba las puntuaciones y no lo adoptó: las evaluaciones humanas indicaban menos fluidez.
Al modelo que calculaba la recompensa le habían pedido evaluar esa fluidez de forma explícita.
In a new benchmark, Visa ran 22 AI models through realistic shopping tasks.
When a run failed, it still ended in a purchase 54% of the time.
That's a purchase decision in a test, not a real checkout.
If your agent ever buys the wrong TV, who eats the return: you or the store?
Vercel AI Gateway added DeepSeek Search + Fetch (Nov 6). Its fallback example sets the order in the system prompt: DeepSeek, then Exa, then Parallel.
Vercel notes the model directs that fallback, so order and retry limits are best effort. I'd log which search tool actually ran.
Neal Khosla calls it "friction collapse": a lot of today's systems were built on assumptions about how humans behave. One example he gives: concert tickets.
When every buyer has an agent that never sleeps, who should get the seat: the fan who waited, or the fastest bot?
Amp Spaces (6 nov): en su ejemplo de revisión de seguridad, el primer pase dividió el código en 16 áreas, comprobó cada hallazgo en otro hilo y dejó un resumen en el espacio para seguimiento humano.
Organización del trabajo — no evaluación independiente de calidad.
Fascinating how quickly many of these AI influencers and evangelists switched sides after a couple bad days. It’s not a bubble but it will be a period of shaking out the weak hands just like in crypto.
Stay diversified and comparmentalized always. The same way they’re not loyal to their customers, you don’t have to be loyal to them. If one provider falls, no problem, isolate that workflow like quarantine and re-route through a different provider. It’s not rocket science.
In OpenAI’s wet-lab benchmark, the same microscope image showed both darker and lighter cell compartments. Astra’s final answer generalized from the darker ones.
OpenAI’s explanation: changing focal depth can reverse the contrast.
Redis Iris Lens (beta): en cada turno compara lo que devolvió cada fuente con lo que entró al prompt.
Los docs ponen el ejemplo: el recuerdo correcto se recuperó… y igual se quedó fuera.
Mistral Managed Agents preview (deployed): a pending tool approval can wait after you close the app — one hour by default.
If nobody answers, that call does not run; the model is told and the turn continues.
SWE-2 Flash: Cognition publica 8 pruebas vs SWE-2. Solo gana numéricamente en una — FrontierCode 2.0, 24,0 frente a 23,8.
En Terminal-Bench 4.0 publican 20,0 para Flash y 27,3 para SWE-2. Cognition lo posiciona más barato y más rápido; la tabla no es un veredicto de producción.
In Revolut’s AI Engineer NYC talk, David Abitbol pairs sandbox rules that limit what coding agents can run with enough telemetry to dig into an incident afterward.
He names the need for logs — not a proof that every incident can be rebuilt.
La guía de Notion para agentes propone guardar tus preferencias y decisiones en una base privada de Memories.
La misma página avisa: MCP actúa con todos tus permisos de Notion.
El agente entra con tu cuenta. Lo privado para tu equipo, tu agente lo puede leer.