No Sheep Uncounted @DreamInCyborg

AI enthusiast gauging the distance before the jump. Joined January 2024

Tweets

113
Followers

27
Following

81
Likes

26

No Sheep Uncounted @DreamInCyborg

10 months ago

@emollick Just found this Isn't CoT still incredibly useful to "bind" the output to a predefined structure or style? That is how I've been using it for some apps. @emollick

0 0 0 44 0

View Details

No Sheep Uncounted @DreamInCyborg

10 months ago

a) Is it possible that a potentially dangerous trait can be expanded, maliciously or otherwise, through apparently innocuous data? b) Is it possible that vectorizing human-generated data will help peer into people's subconscious? Not thrilled with the implications of either.

0 0 0 7 0

View Details

No Sheep Uncounted @DreamInCyborg

10 months ago

2. LLMs transmit this embedded data even unprompted, which is similar to how humans express their subconscious thoughts and preferences through different channels. With that said, I have two big questions/concerns:

1 0 0 9 0

View Details

No Sheep Uncounted @DreamInCyborg

10 months ago

Can seemingly inocuous data be poisoned? Colleague brought up this amazing paper today: arxiv.org/pdf/2507.14805 In a nutshell, when fine-tuning a model with data from another model, behavioral traits are transmitted through unrelated data, such as number series. Thoughts below.

1 0 0 10 0

View Details

No Sheep Uncounted @DreamInCyborg

11 months ago

@iMichaelTen @llama_index This breaks a single vector DB down into multiple, smaller stores on which RAG is performed. Yeah, it's similar, but slightly different.

0 0 1 12 0

View Details

No Sheep Uncounted @DreamInCyborg

12 months ago

@thenicsilver This is very interesting. We have a portfolio of over 100 products and have created our own suite to develop agents, which runs on AutoGen. How should agent supervision be approached under volume? We have soft failures, but we are currently offering limited statistical data.

0 0 0 24 0

View Details

No Sheep Uncounted @DreamInCyborg

12 months ago

@justinskycak I think that is not entirely true; learning also requires understanding (and committing to memory) the logical arguments behind the facts. Naturally, one can say that everything that is not in front of us is in our memory, which, while true, is not useful.

0 0 0 14 0

View Details

No Sheep Uncounted @DreamInCyborg

12 months ago

@hwchase17 @assaf_elovic I have discussed this before with some colleagues: AI apps, in general, have some of the worst UX I've seen around. Be it requiring too much customization upfront, hiding basic options, or just poor decisions in general.

0 0 0 21 0

View Details

No Sheep Uncounted @DreamInCyborg

12 months ago

@andrewallenxo Isn't there a risk that humans would be outcompeted in the tasks they favor? If a human is comparatively unproductive at a task, it is unlikely that would drive employment; that'd be a hobby.

1 0 0 13 0

View Details

No Sheep Uncounted @DreamInCyborg

a year ago

@illyism Isn't this worse than using a vector database in almost every conceivable way?

0 0 0 8 0

View Details

No Sheep Uncounted @DreamInCyborg

a year ago

@emollick We have developed a multi-agent system to handle workflows based on AutoGen AgentChat. I have found that a lot of the hardest flows/prompts end up working as expected once you understand the models, data visibility, etc. Prompting IS a skill when the prompt is complex enough.

0 0 0 18 0

View Details

No Sheep Uncounted @DreamInCyborg

a year ago

Will AI make top performers multiply their productivity, or will it make average performers perform at a top level?

Alex Albert @alexalbert__

a year ago

Claude raises the waterline for competence. It handles 99% of general discrete tasks better than the average person - the kind of work that used to require basic professional training but not deep expertise. This creates an interesting split: routine generalist work (basic

58 123 1K 139K 538

0 0 0 12 0

View Details

No Sheep Uncounted @DreamInCyborg

a year ago

@Sumanth_077 Very interesting! I am currently interested in multi-agent systems; what is your take on these vs single-agent?

0 0 0 143 0

View Details

No Sheep Uncounted @DreamInCyborg

a year ago

@TheGeorgePu I agree with this. If there's something AI is definitely going to do, it is to drive the cost of certain types of software, like enterprise solutions, down to near zero. Brands will still survive, as well as apps that thrive on higher user counts, like social media.

0 0 0 8 0

View Details

No Sheep Uncounted @DreamInCyborg

a year ago

@bindureddy @emollick The key is to build a modular system that uses implicit intent capture to route the requests down the right AI pipeline. I've been working for two months on such a system: specialized agents do wonders.

0 0 0 4 0

View Details

No Sheep Uncounted @DreamInCyborg

a year ago

@ChenAIResearch @JeremyNguyenPhD @voiceflow We use VoiceFlow for an 80+ catalog of products. We mix its built-in capabilities with our own tools and agentic systems. Voiceflow is good but a bit constrained; we have tried other products like Amelia or Forethought, which are more flexible but less predictable.

0 1 3 60 1

View Details

No Sheep Uncounted @DreamInCyborg

a year ago

@ankurnagpal In nearly 100 products we have achieved an overall rate of 65% rate, with some products going as far as 90%. The first 50% is easy enough using RAG, after that, the key is to have specific tooling. My favorite? An AI-ops system we built for ruby-as-a-service.

0 0 1 41 0

View Details

No Sheep Uncounted @DreamInCyborg

a year ago

@rohanpaul_ai My experience with using AI for workflows is that, as the number of steps increases, AI also becomes more rigid in its approach and is less able to deviate (within the space provided). I think a "checklist" approach would work great for these things.