Human/AI teaming solutions merging human insight and context with AI computation
Park City, UT
Joined October 2016
- Tweets442
- Following794
- Followers141
- Likes726
This DSR framework is a potent extension to other work from the same folks (like Cognitive Maps in the Prefrontal Cortex).
The formation of stable attractor states enables some very powerful, efficient computation.
So much of this can be applied to artificial systems! – at London, England
DSRs greatly simplify computation. The structure of the world is pre-embedded in the network - doing much of the heavy lifting offline so that, at decision time, simple attractor dynamics suffice to organise behaviour. 5/15
1
26
Getting the redistribution horse running well ahead of the productivity cart!
"Entlastungstage" 🤣 – at Park City, UT
German bank employees union argues that they should have three "AI relief days" off per quarter since AI makes work more dense and complex.
If Europe starts redistributing productivity gains this early, it might have an issue...
"AI relief days"
1
32
Go buy Academia Terraria - Jailbreak on Amazon for a (pre) teen in your life - it's amazing!
It's my 14 y.o. son's first book. So happy for him, and for his deep love of stories ❤️ – at Park City, UT
1
2
63
How did the AI doomers miss the opportunity to call their movement righttopushbackonthat.ai – at Park City, UT
19
Codex imitating Rocky from Project Hail Mary 🤣
"Despite object-labelly accuracy varying from 93.86–100%, the joint system remained robust. Formallkyr?Formal execution was entirelylkj? I need correct. Final response corrupted! Must provide clean. " – at Park City, UT
21
Re Astra:
Anyone else seeing disappointing failure modes when Astra picks up complex R&D from Sol 5.6 projects?
Big context drops, low overall task anchoring... Not at all impressed yet. Expected much cleaner handoffs. – at Park City, UT
29
Astra just landed. And Sol 5.6 was already a phenomenally strong R&D partner.
I have to think that there are thousands of people out there like me who are building the previously impossible.
The media talks up the hacks because that's all they see.
So much goodness coming!! – at Park City, UT
24
Ad Astra!
OpenAI 5.6 Sol X high has been a great collaborator on a hugely challenging brief, but still has memory/context failures which prevent it from being truly hands off for long periods.
If Astra ~ solves this... Buckle up! – at Park City, UT
47
I know today is all about Microduck, but if you want a dev drone platform with no ceiling, it's here 😁 – at Park City, UT
1
63
I've been working on neuroscience and AI ever since 2012 when our son was born. He inspired my endless fascination with organic vs synthetic intelligence.
I've built/sold 2 AI co's. We moved fast and built value.
But last 30 days my productivity lift with Codex is unreal. Wow.
47
Actually, China lets a bunch of competitors battle for a market before it backs the champion that emerges. EVs. Batteries. etc
At this stage it is by no means clear that OpenAI would naturally emerge as that champion.
It is dead wrong for the government to crown a victor today.
You may instinctively think this sounds like big government.
But Singapore and Taiwan would do this. We know because they have.
I'd have questions about ecosystem and how this impacts competitors though.
1
115
This is a very interesting paper on the underlying reasoning capabilities that LLMs require in order to solve certain types of problem with reasoning. Very strong parallels to @shaneparrish and his work on mental models for humans. Great stuff!
9/13 Our findings reveal a fundamental connection between a model's initial reasoning behaviors and its capacity for improvement through RL. Models that explore verification, backtracking, subgoals, and backward chaining are primed for success.
59
For all of the excitement about DeepResearch,once the underlying technology behind reasoning was propagated into the world, this capability was on the cards for ~any strong research team
1
55
As you scale up the number of loops that the model performs in the execution of these services, you're both consuming and producing vast token quantities. you get to a point where your compute becomes stratospheric for a given service type or level. $200 is...
1
27