Saoirse

Personalized Deep Research in Percolate And why agent memory plus AI search may trump reasoning

open.substack.com/pub/percolat...

Inception Labs We are leveraging diffusion technology to develop a new generation of LLMs. Our dLLMs are much faster and more efficient than traditional auto-regressive LLMs. And diffusion models are more accurate, ...

Parallel Text Generation Now Live 🚀
www.inceptionlabs.ai/news

thinkingmachines.ai

"Instead of focusing solely on making fully autonomous AI systems, we are excited to build multimodal systems that work with people collaboratively,"

Detecting hallucinations in large language models using semantic entropy - Nature Hallucinations (confabulations) in large language model systems can be tackled by measuring uncertainty about the meanings of generated responses rather than the text itself to improve question-a...

Im a sucker for information and entropy - and im glad folks are starting to be apologetic about using the world Hallucinations as we hopefully adopt confabulations instead - www.nature.com/articles/s41...

No LLM is an Island No matter how close we get to AGI, we should always keep data integration in mind

I elaborated on this here percolationlabs.substack.com/p/no-llm-is-...

im confused by what many consider a dichotomy between RAG and large context as if RAG is merely a stopgap til LLMs become omniscient. RAGs poor connotations due to its genesis in vector retrieval belie that its the broader need for integration; search, personalization, tool use. no LLM is an island

Fully autonomous AI agents should not be developed.... semi-autonomous systems, which retain some level of human control, offer a more favorable risk-benefit profile, depending on the degree of autonomy, the complexity of tasks assigned, and the nature of human involvement. arxiv.org/pdf/2502.02649

Percolate - Getting Started YouTube video by Percolation Labs

I spent three weeks building a "data tier agentic orchestrator" prototype from scratch and 1 week creating a 10 minute video about it. Making videos is hard! Hopefully it gets easier.
www.youtube.com/watch?v=SWUT...

While i think this is very useful, there is something about this analytical deconstruction that loses something. The most fun ive had this year understanding cognition is still the philosophical discussions in _The Matter With Things_ (McGilchrist)

🦸🏻#10: Does Present-Day GenAI Actually Reason? A Blog post by Ksenia Se on Hugging Face

Do LLMs Reason? - good read
huggingface.co/blog/Ksenias...

🦸🏻#10: Does Present-Day GenAI Actually Reason? A Blog post by Ksenia Se on Hugging Face

Read more in our article. We discuss if models really "think" or not, where AI already excels, and what its limitations are -> huggingface.co/blog/Ksenias...

In the limit, agentic workflows will be built from data, not code Why we are betting on an agentic data tier

Check out our thoughts in this post open.substack.com/pub/percolat...

Are agents code or data?

AI / Critical thinking

www.microsoft.com/en-us/resear...

This is fascinating (edelman.com). Can we say both developing world & also east are more trusting of AI than west & developed countries? Im maybe not surprised Ireland is so low in trust. Trusting AI may be considered a sort of figairey. The chart also compares trust in US vs China. Spicy.

Creating at the speed of thought How large language models will mostly just reduce our tolerance for typing

Generative AI is just better UX...

percolationlabs.substack.com/p/creating-a...

People sometimes forget that generative AI is more like the thing on the left than the thing on the right

What We're Fighting For Soundtrack: Bad Religion — The Resist Stance A great deal of what I write feels like narrating the end of the world — watching as the growth-at-all-costs, hyper-financialized Rot Economy seemingly ta...

Newsletter: The tech industry has turned on the consumer, using abusive, manipulative software as a means of extracting wealth from every corner of our lives. Both society and the tech media must fight back, as I believe many of us deeply love our tech.

www.wheresyoured.at/what-were-fighting-for/

GitHub - Percolation-Labs/percolate: An agentic orchestrator *inside* a relational-vector-graph/key-value database! An agentic orchestrator *inside* a relational-vector-graph/key-value database! - Percolation-Labs/percolate

Three weeks into building Percolate, an agentic orchestrator in the data tier using a relational-vector-graph-key database. I oscillate between excited and wondering if its a crazy idea :)
github.com/Percolation-...

Series: Animals Wearing Human Clothes: Penguins Wearing Sweaters #penguins #ai #aiphoto #cute #love #fashion #beautiful #sweaters #aiphotography #portrait #aiphotographer #davinciai #artificialintelligence #machinelearning #aiart #digitalart #technology #art #midjourney #aiartcommunity #datascience

Researchers created an open rival to OpenAI's o1 'reasoning' model for under $50 | TechCrunch AI researchers at Stanford and the University of Washington were able to train an AI "reasoning" model for under $50 in cloud compute credits, according

AI researchers at Stanford and the University of Washington were able to train an AI “reasoning” model for under $50 in cloud compute credits, according to a new research paper released last Friday.

Generative retrieval and the concept of percolation in statistical physics Why is engineering RAG like the idea of percolation in statistical physics?

percolationlabs.substack.com/p/generative...

Setting up K8s on a dedicated Hetzner “Robot” Server I am documenting my experience setting up for the first time with Hetzner a dedicated server provider that has a great service at great…

thats where my thinking has been going .... medium.com/percolation-...

Setting up K8s on a dedicated Hetzner “Robot” Server I am documenting my experience setting up for the first time with Hetzner a dedicated server provider that has a great service at great…

While I use Substack for less technical articles, I tend to write more technical articles about Percolate on Medium. my latest medium article is about P8s + K8s + Hetzner. In the past, AWS, GCP or DO have been my default choice for this sort of thing - maybe not in future
medium.com/percolation-...

GitHub - Percolation-Labs/percolate Contribute to Percolation-Labs/percolate development by creating an account on GitHub.

github.com/Percolation-...

If you looked at the codebase you might have seen that we use best in class wasm serverless architecture that will not only be green but save $$ by only processing cpu/memory when an action is taken. And, to your great point we are using the universal Postgres db as our first step towards spannerdb