Releasing the KAOS Framework

Super excited to release the K8s Agent Orchestration Framework (KAOS) to help manage distributed agentic systems at scale 🚀 The KAOS Framework addresses some of the pains of taking multi-agent / multi-tool / multi-model systems to hundreds or thousands of services! It started as an experiment to build agentic copilots, and has progressed as a fun endevour building distributed systems for A2A, MCP Servers, and model inference! The initial release comes with a few key features including: 1) a golang control plane to manage Agentic CRDs; 2) a python data plane that implements a2a, memory, tool / model mgmt; 3) a React UI for CRUD+debugging, and; 4) a robust CI/CD setup with KIND/pytest/ginko/etc. I have to say I am impressed on the level of abstraction that is possible to reach with agentic copilots when covering frameworks and domains I have experience in, a blog post will follow on this topic specifically! For the meantime do check out the repo, docs and examples to try it out - if you have any feedback or run into issues please do submit an issue!

FLUX: Interactive Visual Intelligence

Black Forest Labs has released another impressive text-to-image model with impressive low-latency/low-compute: It seems that now we're on a trend where the next big milestone is ultra-low-latency models designed for interactive real-time applications, and I also believe this is a clear path for success. This is particularly interesting as it claims sub-second end-to-end text-to-image and image editing while running on consumer GPUs (the 4B variant fits in ~13GB VRAM). The key product idea is a single unified architecture that supports text-to-image, image-to-image editing, and multi-reference generation, offered in distilled "fast" variants. The best thing is that it's actually Apache License which means it's actually Open Source, so kudos to the team for opening this to the masses.

Counterfactuals for RecSys

Anyone that has run A/B tests for RecSys evaluation knows the challenges, which is why it was super interesting to come across Eugene Yan's Counterfactual Evaluation for Recommendation Systems: Counterfactual (ie off-policy) evaluation basically helps us estimate "what would have happened if". Namely this works using logs from the current policy, apparently which is most commonly via Inverse Propensity Scoring which reweights each logged reward by the ratio of the new policy's probability of the logged recommendation to the production policy's probability. There are some downsides as apparently IPS can break with insufficient support (aka zero propensities for actions the old policy never took) and can have high variance when probability ratios explode. There are solutions to this problem such as stabilizers like the ones presented as Clipped IPS  and Self-Normalized IPS - which is quite interesting, definitely a great resource to an more mature field (but still unresolved) in the MLOps ecosystem.

Universal Commerce Protocol

Google has co-released the Universal Commerce Protocol with top e-commerce brands like Shopify, Zalando, Etsy and Wallmart to enable Agentic Shopping! This is quite an interesting trend that likely will reach other intdustries; particularly in e-commerce this is currently changing the way individuals interact with their day to day purchases - I have found myself using LLMs more often than Web Search to rate and sort products. This standard defines how agents discover a merchant through standardised APIs which allow negotiation, and then transact using stable core checkout primitives augmented by independently versioned extension schemas. It bakes in a human-in-the-loop through an explicit checkout state machine and continue_url handoff, which ensures that shopping experience still remains seamless through an embedded checkout protocol with bidirectional JSON-RPC messaging and delegated UI affordances. It is quite interesting to see that the standads define even payment flows, which are treated as a two-sided negotiation where merchants dynamically advertise allowable payment "handlers" for the specific cart/context and agents execute the corresponding provider-authored specs while keeping raw credentials out of the platform to reduce PCI scope. It really feels like we are living in the future sometimes!

DuckDB for Everything Data

Data processing is the quiet force multiplier behind every successful ML system, and it seems that DuckDB is taking quite some of that market share: DuckDB is an excellent default for ML/data engineering pipelines because it combines the simplicity of an in-process database with OLAP-grade performance for joins and aggregations, often far faster than OLTP engines for analytics workloads. It's quite a nice single-binary pip install and allows for rapid local iteration (plus CI/testing). I have used it in a few projects and it has been quite impressive particularly the CSV parsing (+ has similar level quality for other parsers on Parquet/JSON/etc, as well as directly from disk/S3/HTTP). There are quite a lot of growing integrations between old guard tech like Postgres and DuckDB, as well as approaches embedding DuckDB, this is really an interesting space to keep an eye!

The Institute for Ethical AI & Machine Learning is a European research centre that carries out world-class research into responsible machine learning.

✉️ Email, 🐦 Twitter, 💼 Linkedin