ro

anyone else just using gpt-5.6 sol for everything?

@ben_ai_eng 100%

Removing the 5-hour Codex window has honestly been a bigger qol improvement than i expected. I always end up working well beyond 5 hours anyway, so tracking both the 5-hour window and weekly usage just created unnecessary context switching. Now i can just focus on problem solving. Less than 48 hours in, it's already reduced my day-to-day stress during a critical production launch. Hopefully this stays in place. Nice change, @thsottiaux @OpenAIDevs 🤝

i think the market is vastly overestimating demand for every enterprise training their own models. frontier lab foundation models will continue to improve through transfer learning, while the cost and operational burden of custom training remains high. see gpt-5.6 sol

@fredrikalindh how do you assess / are you watching the whole time?

i wrote a post on how cursor does hiring 6 months back, but a lot has changed bc of the latest models now onsite is just 5h to ship a full product from scratch here's what the people who pass do differently:

@ziv_ravid the models are only going to only keep getting bigger

I wonder whether, in retrospect, Anthropic thinks it made a mistake in developing such a large model. It made more sense in February, when there was no model near its performance level, but now, their only option is to keep subsidizing it through the paid plan

@catboosted this ^

I’d love to double down on the meme but this is very sophisticated trolling Researchers do not want your stupid random benchmarks or data. They only care about hillclimbing economically valuable work The top N vendors sell capability lift, not trash.

Jevons Paradox is about to hit AI harder than almost any industry we have seen before. People once thought faster internet would simply let us load the same websites more quickly. That was not even close to what happened. Faster connections created video streaming, cloud software, online gaming, video calls, social media, and entire businesses that could not exist on slow internet. Every increase in speed created new reasons to use more bandwidth. AI will work the same way. Today, we mostly use models for chat, coding, search, writing, and a few business workflows. But once intelligence becomes cheap enough, fast enough, and reliable enough, it will be built into every process that involves a decision. The biggest AI workloads probably do not exist yet. They are waiting for the cost of intelligence to fall. It will create millions of new tasks that are currently too slow, too expensive, or simply impossible.

@Hesamation impressive today, but frontier models are scaling faster than local hardware memory.

Apple’s M7 Ultra chip rumored to be a local AI workhorse as it supports 1.5TB of RAM. for reference, that’s enough to run Kimi or GLM models locally.

Today we share the worldview behind our mission. Human values don't average out. Local knowledge can't be centralized. The good future has many AIs, raised in different places, shaped by the people they serve, disagreeing with each other the way we do. thinkingmachines.ai/blog/the-futur…

@alexatallah shouldn’t we believe in the generalized representations that comes from transfer learning though?

Very important perspective that will grow over the course of the year: Enterprises preserving their IP, improving their AI neurodiversity, & creating hill-climbing evals to build "veteran" agents that beat generalist models. Proprietary evals will finally become a thing!

@nikunj outbound sales is our favorite activity

There’s nothing more humbling than outbound sales. Like any skill, it’s learnable. But man, when you see the really great ones do it, you can’t help but walk away a bit envious. Going to be even more important as time progresses!

@xdotli why?

for everyone who's coming to sf and uses laptops buy this you will thank me later

@aashaysanghvi_ all about the narrative

Lord, give me the confidence to assign a long term multiple to revenue from labs

@fredrikalindh @steipete in our interview process we screen for how agent native the candidates are! also, super critical to give unlimited agent access to get the best candidates

if i were looking for a job today i’d prioritize working with people who are agent maxxing and have unlimited tokens. you want to position yourself to become extremely good at working with agents and the only way to do that is to do it a lot

When we were starting @mercor_ai, most of the top investors in Silicon Valley told us that the AI data market would collapse. They had no understanding of the market, but advised us that we should pivot the business to another industry. The most dangerous thing a founder can do is be too deferential to investors with less context. Nvidia, OpenAI, Anthropic, SpaceX, Meta, and most of the legendary companies started as deeply contrarian. If any of the founders had listened to investor consensus, those businesses would not exist.

@BarrAlexandra this is very true lol

someone at a party last night said VCs will ask you what childhood trauma you’re recovering from, and if you say you don’t have any, they won’t fund you

@h100envy @ycombinator TIL there are fine-tuning boot camps 😂

Ex-NVIDIA engineer who built Unsloth explained RL, kernels, reasoning, quantization, and agents in 2 hours 42 minutes - better than $5000 fine-tuning bootcamps. pick the base model -> write triton kernels for 2x faster fine-tune -> quantize to 4-bit -> run GRPO/DPO -> ship a reasoning model on your single GPU. That loop is why Unsloth is the default way to fine-tune Llama, Qwen, Gemma, and Phi on hardware you already own. Unsloth + Triton kernels + 4-bit quantization + GRPO/DPO + single-GPU fine-tuning - that's the stack. Watch and save it, then fine-tune your first model tonight.

Today we’re releasing APK Arena: a frontier benchmark for evaluating AI agents on phone-use. We ran GPT 5.6 for a side by side comparison where it outperformed other models across the board and is currently leading on APK Arena. 50+ levels: taps, vision, memory, games, real tasks. Here's how it did vs every other frontier model 🧵

@jeff_weinstein @stripe We’d be interested cc @gabrieljfong

Looking for a few SaaS startups that: - use @stripe for subscriptions, - don't yet run [Google, Meta, etc] ads, - want to run their first ads, - and would be willing to intensely give us feedback on a dashboard for analyzing the impact of ad spend. dm or jweinstein@stripe.com

A surprising amount of development work is really cognitive science disguised as engineering.