alex funk

the dream machine i made for the yc hackathon today. thanks @AgentPhoneHQ , @supermemory , @GoogleDeepMind and ofc @ycombinator #ychackathon

if you don't know michael sugrue - you should. he might be the greatest philosophy lecturer of all time. his videos got me through my philosophy degree a must watch. youtube.com/watch?v=RFNLzJ…

The Attention Schema: To control and optimize this deep attention, the brain creates a cartoonish, simplified model of it: the attention schema. When the brain reads this internal data model, it concludes: "I have an invisible, subjective property called 'awareness' directed at that object."

In terms of theory of mind we are getting quite close to technical definitions of consciousness. Especially once you start including loops.

Nature vs nurture Is someone born with high intelligence? Model weights are the “nature” half of the coin - set in stone, unchanging. Context layers (obsidian, Toto, gbrain, etc) are the “nurture” side. Based on experiential feedback they store an agents experiential memory

@Bhavani_00007 There’s definitely a bubble with this style of thinking. Software is “free” now but my barber does not know anything about building an application - there is still a very real market for targeted solutions

Everyone says: "SaaS is dead" "Software is free everywhere" Cool!!!!! Then go build your own Slack, GitHub, Notion, Zoom, 1Password, Jira, and Rippling this weekend. What's stopping you? Seriously, I want to know.

@DhravyaShah @kushthaker03 Any inspiration from the Toto dream machine 👀

I really do not trust agent “score” what the fuck does 7.6 mean? Agent based scoring seems incredibly arbitrary unless given rigorous guidelines

@levie That is the core problem toto.tech is solving. Routing your token spend to the model most appropriate for the work. You don’t need rocket fuel in a Prius - you don’t need opus to write LinkedIn messages but the end user has to manually choose the model they use rn.

Token costs will become a dominant topic in enterprises going forward with AI. Just got out of a dinner with many Fortune 500 enterprise CIOs and this was the most heated topic. A mix of strategies are being employed, but basically no one feels like they have the right solution. A mix of: figuring out how to prioritize workloads to different models, giving out access to better or worse agents by user type, setting different spend caps by team, having teams justify AI by their use-case, and some just having unfettered access. Everyone is trying to figure out a semi/predictable model right now in a world where the underlying tech and cost models are constantly evolving.

@garrytan My thesis is local models can be used for effective routing tools to frontier labs ultimately reducing token spend dramatically

Qwen2.5‑7B Instruct apparently is GPT-3.5-turbo level

@adrielyong Guinness is good for you

my guilty pleasure is beer but beer is also the worst for weight gain / gut i wonder if theres a concoction of supplements that can neutralize the negative effects in my tummy from my nice cold beer

The hardware bottleneck is manufactured to prevent you from building your own LLMs

@kenwuuuu I’m using a lanchain + pi combo.

so what is the current industry standard to build agents? do people still use langchain/langgraph?

does anyone else build their own testbenches? small, self contained environments with rich sliders and detail?

@hthieblot @JoshuaKushner

I don’t find this to be a badge of honor. If you live like this I imagine the UX of your product feels quite similar to this. How you do one thing is how you do everything

@paulitics_ @garrytan if you can route work to an array of fine tuned modals for a set of discrete types of tasks, you can reduce your price/mTOK dramatically without a huge drop off in quality

@garrytan why fine tune though? surely you expect the next frontier model to outperform your fine tuned version anyways? I guess cost?

Thinking Machines is impressive. In a couple hours I just fine tuned my own Qwen3.5-397B model this afternoon. Fast usable multimodal is also going to enable very mind-blowing personal AI.

@gdb It’s one of the core reasons I use it everyday as my primary harness.

under appreciated that codex is open source

@MegaBasedChad It’s the feeling of the sublime I think

Why is it that being around mountains feels so great? I was always told that the Plains are Humanity's natural habitat. Yet mountains produce a universal feeling of splendor, much more than the plain.

steer is inferior to claude code's interruptions with new messages. steer is like a 90 degree turn at times.

@mil000 Were you the guy at the claw machine yesterday

It’s a D tier accelerator that blows half of their LP money on just the warehouse. It’s not “early hacker Silicon Valley”