This week Miguel Fierro, a former Microsoft principal researcher who not too long ago based his personal firm, RecoMind, joined knowledge and AI evangelist Christina Stathopoulos to speak concerning the state of advice programs. Christina additionally ran by the newest AI information she’s been watching, from Anthropic’s continued rise to accountable AI, bulletins from Google’s I/O 2026 convention, and (persevering with the dialogue from final week) the rising backlash in opposition to tokenmaxxing as a productiveness metric. Listed below are three takeaways from the dialog.
Suggestion programs are a much bigger deal than most corporations understand
Miguel has spent the higher a part of a decade constructing suggestion programs for enterprise clients at Microsoft, and he thinks most corporations are leaving rather a lot on the desk by not paying nearer consideration to suggestions. Amazon generates roughly 35% of its income by suggestions. Netflix attributes 75% of content material consumption to them. Greatest Purchase credit suggestions with 24% of income. TikTok’s complete consumer expertise is a suggestion engine. And but many massive retailers he labored with at Microsoft weren’t investing critically within the space, actually because they weren’t monitoring the worth it was producing.
The hole between the highest tier and everybody else is broad and getting wider. Essentially the most superior programs at the moment deal with consumer conduct as a sequence prediction downside, just like how massive language fashions predict the following token. Reasonably than simply encoding clicks, they encode all consumer actions into embeddings, run sequences by these representations, and use large 1.5 trillion-parameter fashions to foretell what a consumer will need subsequent. That’s not one thing a mid-tier retailer can replicate at the moment, but it surely indicators the place the sector is heading.
Even should you don’t work in a prime well-resourced firm, it is best to nonetheless take note of the convergence of search and suggestions right into a single personalised retrieval layer and the early utility of basis fashions to suggestion issues. Netflix has constructed what Miquel described because the solely printed basis mannequin on this area; Meta is rumored to be creating one as properly. The barrier is knowledge, notably for smaller organizations. In contrast to textual content, behavioral interplay knowledge isn’t publicly obtainable, so constructing at that scale requires each proprietary datasets and critical compute.
If you wish to get your palms on state-of-the-art implementations, together with data graph-based approaches, with out ranging from scratch, Miguel steered the open supply Recommenders library, initially developed at Microsoft and now housed underneath the Linux Basis, as a sensible entry level.
The agent hype has a recommender-shaped gap in it
Miguel drew a distinction between true gross sales brokers and what most corporations provide at the moment, that are normally simply conversational brokers. A conversational agent responds to what you say. An agentic gross sales system understands a buyer, anticipates what they need, and surfaces the appropriate product or provide on the proper second—and that requires a suggestion system baked in.
In case your “agent” is a chatbot with entry to a data base, it’s not doing suggestion. Suggestion programs want coaching knowledge, a retrieval layer, and a personalization mannequin, none of which you get totally free from a basis mannequin API. A language mannequin can reply questions on a product catalog, however it will possibly’t provide up personalised suggestions until it additionally has a mannequin of the client’s preferences, historical past, and sure subsequent motion. Most corporations don’t have the infrastructure in place to make that potential. . .but.
The accountable AI dialog has left the analysis neighborhood
What’s notable concerning the accountable AI dialog proper now’s the vary of establishments providing their perspective. Anthropic, alongside saying a funding spherical pushing its valuation towards $1 trillion, urged a world pause on AI growth tied to the danger of recursive self-improvement: programs that may design and develop their very own successors. The Way forward for Life Institute printed The Higher Path for AI, a framework arguing for functionality growth oriented towards human profit somewhat than human substitute. And the pope issued a formal encyclical centered on AI and the frequent good.
None of those establishments is making the identical argument, however the convergence of their consideration issues. Accountable AI was once a specialised dialog occurring largely inside analysis labs and a small set of coverage organizations. It’s now a subject the place main AI corporations, spiritual establishments, and civil society teams are all staking out public positions in the identical information cycle.
For the technical neighborhood, this creates each strain and alternative. “We’re fascinated about security” is now not a enough posture; exterior scrutiny is intensifying from instructions that don’t share the sector’s assumptions or vocabulary. However the broader dialog creates actual demand for practitioners who can translate between what accountable AI really requires in follow and what policymakers, executives, and establishments are attempting to determine. That translation work is more and more the place the sector wants folks.
What’s subsequent
Be a part of us Monday morning for the subsequent episode of This Week in AI, the place YK Sugi and John Lindquist will break down the huge structural and monetary shifts reshaping the expertise trade. (They’ll additionally chat concerning the latest launch of Claude Fable 5.) And on July 23, Christina will probably be internet hosting the AI Superstream on AI harnesses, a four-hour occasion centered on agentic AI and the frameworks practitioners want to maneuver from fashions to brokers. Each are free to attend. Register now to save lots of your seat.
For deeper studying on matters coated this week, Christina advisable three titles obtainable on the O’Reilly studying platform: Arms-On LLM Serving and Optimization, Arms-On RAG for Manufacturing, and Massive Language Fashions: The Laborious Elements. Not a member? Join a free 10-day trial to verify them out.
We’ll proceed to publish our takeaways right here on Radar every Friday and share full episodes on YouTube, Spotify, Apple, or wherever you get your podcasts.
