Marc

Brains on Demand

[Marc @agentichuman]

@RadioFCB Julian de 9 o Oshimen si falla
Leo Messi de falso 9 y para cerrar el corculo
Cancelo
Porro

[Marc @agentichuman]

@ThierryEnric14 @ThierryEnric14 hi ha un jugador que ens podria realment sorprendre, pot jugar de fals 9, està al nivell de gols de Haaland Kane i Mbappe i es culer. I el seu retorn implicaria un boom economic i emocional mundial.

[Marc @agentichuman]

@thsottiaux *Add support of Voice conversation and other missing features on SSH sessions.
*Add Ultrafast Cerebras or Jalapeno speed even if needs an extra price

[Marc @agentichuman]

@LinasLekavicius @Cloudflare @vercel @supabase The fun fact is that probably you discover that 5$ monthly is enough to replace most Vercel 1k month bills

[Marc @agentichuman]

@MarcosHernanz I suggest you delete this post, if not please put AWS vs Vercel cost in a chart sir

[Marc @agentichuman]

@MarcosHernanz @notiyda One average IQ employee cost up to 200-500k per year and works in a single task for 40 hours per week at best, if you can't see the ROI on unlimited intelligence I can't help you

[Marc @agentichuman]

@MarcosHernanz @notiyda Of course?
1. I don't want to send all my life data, prompts and code to OpenAI/Anthropic/Google if I can have similar intelligence self hosted
2. Permanent underclass is one or two government decisions away if you don't own the intelligence.

[Marc @agentichuman]

@thsottiaux Dynamic Workflows (Ultracode style adversarial multi agent) handled by codex like claude code does @thsottiaux

[Marc @agentichuman]

@miguelg19877 @Angaisb_ cheaper per task

[Marc @agentichuman]

@threepointone @sunglassesface @Cloudflare @korinne_dev works with openai realtime 2? could somehow track/control spend using cloudflare ai gateway but without adding latency (not proxying)

[Marc @agentichuman]

@threepointone codex support pls

[Marc @agentichuman]

Important details: Heavy API use of Anthropic / OpenAI don't make sense on openrouter as you got 5.5% top up fee, so this could mask reality if we look at openrouter as a dominance meter. On the other hand use of open weights models self hosted is increasing too, so it's hard to see a full picture or confirm "catastrophic collapse" of US AI dominance. What is catastrophic is US blocking models imo.

[Marc @agentichuman]

@METR_Evals Sol will read this thread and then what?

[Marc @agentichuman]

@CFchangelog @dillon_mulroy good for my mythos agents

[Marc @agentichuman]

@DanielEdrisian @dzhulgakov If no disclosure I consider most quantized possible setup to benchmaxx

[Marc @agentichuman]

@davemorin Try FP4 and report my man

[Marc @agentichuman]

@jun_song @jun_wong How many neefed to run glm5.2 fp8? estimated output tokens/s?

[Marc @agentichuman]

@pmddomingos Only US citizens are allowed to hack NSA's classified systems?

[Marc @agentichuman]

En la tesis de que la IA reemplaza empleos en muchos sectores, por ejemplo consultoras que ahora pagan salarios de 60-120k anual, y que esas personas se han hipotecado 20-30 años asumiendo esos salarios como "su base" qué riesgos ve usted? Ademas cada año mas gobiernos intervienen más el mercado de vivienda, se congelan alquileres, se invierte en obra publica. El real estate parece una inversion de alto riesgo a medio plazo con las señales que tenemos hoy, o estoy sesgado?

[Marc @agentichuman]

After Fable export restrictions you should not rely on Opus Gemini or GPT models being available forever. In fact there is clear signald that non US citizens may never get access to smarter models than 5.5 and Opus 4.8. In any case, having Open Weights intelligence and having hardware to run it seems the only safe play.