June 17, 2026

Dear Reader,

This week we continue celebrating the 7th anniversary of the Data Science Briefing. Over the years many things have changed. We switched platforms behinds the scenes a couple of times, the format evolved, and, link rot took its toll. If you’ve recently ventured into the list of past issues in https://data4sci.com/newsletter you might have noticed that many of the older issues were no longer available, but no longer.

Over the past few weeks, we’ve gone on a deep Internet spelunking expedition to reconstitute the past of this humble venue and created local copies of all the past issues. You can now checkout the very first issue, the first year anniversary, the 100th, 200th or even 300th issue, or any other issue you’re curious about.

Over the past 322 issues, we covered 2,508 links, 2,501 papers, 311 videos, and reviewed 65 books. Not bad for something that got started almost on a whim :o)

The biggest models of 2026 owe more to old math than the launch posts admit. Seventy years of AI research keep pointing to one pattern. General methods that throw more computing power at a problem win, and they win by a wide margin. Hand-built human knowledge looks clever at first, then it stalls. The pattern repeats in the newest joint-embedding models. Their core trick traces straight back to a 1936 statistics paper on canonical correlation analysis, a 90-year-old idea for finding a shared signal across two sets of data. JEPA does the same thing in a non-linear form: predict one view of the data from another, in embedding space. The credit for these systems runs back further than most readers expect.

The race to build one smartest model is no longer the whole story. A new routing method fans a prompt out to a panel of models, then lets a judge merge the answers. On 100 deep research tasks, a fused pair scored 69.0%, above the 65.3% of the best single model. A trio of cheaper models came within a point of that top model, at half the cost. A mix of models beats sheer size. The same shift drives the rise of agent harness engineering. The model is one input. The prompts, tools, context rules, and feedback loops around it do the rest of the work. A capable model inside a strong harness beats a stronger model inside a weak one. But an early look at the newest model cuts the other way. It ran for up to a dozen hours on multi-page specs and beat every public model in the tests. The strange part was how little the human did, and how little control the user kept over the model’s choices. That is the open question for practitioners now: not raw power, but how much control we keep.

Retrieval is splitting into two camps. One camp keeps building richer structure. The method that kicked off graph-based RAG turns a document set into a knowledge graph, groups the entities into communities, and pre-writes a summary for each. Ask a broad question like “what are the main themes here,” and it map-reduces those summaries into one answer, even across a million tokens. Standard vector RAG can’t do that. A newer system fixes a weak spot in that pipeline. Graphs built fragment by fragment turn out messy and self-contradictory, so it puts a team of agents on shared memory to build one consistent graph, then retrieves over it. The other camp asks if any of this is needed. One study pitted plain grep against vector search across several agent setups on long chat histories. Grep won more often for pulling back exact names, paths, and error strings. The sharper finding: the harness around the tool moved the score more than the search method did.

Run these agents and the bill arrives fast. A recent study traced eight frontier models through coding tasks and found the work burns about 1,000 times the tokens of a one-shot coding reply. Input drives the cost, not output. The agent re-reads its whole history before every step. Worse, the spend is erratic. The same task can vary 30 times in tokens, and paying more buys no extra accuracy past a point. The models cannot even guess their own cost. A second study broke down a multi-agent coding pipeline and put a number on where the money goes. Roughly 59% of tokens went to the review-and-revise loop, not to writing the first draft. So the cost lives in checking the work. A third team hands the cleanup to the agent itself. Their loop mines an agent’s own failures, proposes small harness edits, and keeps only the ones that pass a regression test. On a terminal benchmark, one model’s pass rate climbed from 40.5% to 61.9%, with no fine-tuning and no stronger model coaching it.

Step back from agents, and structure still decides the outcome. One piece argues the internet’s wiring, not just its content, strains democracy. The rollout of US broadband 20 years ago varied by state cable-laying rules, a near-accidental experiment. Areas that got broadband first saw political polarization rise. A modeling study makes the same point with pathogens. The 2026 World Cup will route 1 to 5 million fans into 11 US cities over five weeks, and the match schedule decides who lands where. Map flight paths onto that schedule, and dengue from Brazil tops the risk at most venues. Atlanta breaks the pattern, with malaria arriving on direct flights from West and Central Africa. Different harms, same lesson. The shape of the network sets the danger.

Our current book recommendation is “Building Applications with AI Agents” by M. Albada. In this week’s video, we have a talk by E. Eifrem titled GraphRAG: The Marriage of Knowledge Graphs and RAG.

Data shows that the best way for a newsletter to grow is by word of mouth, so if you think one of your friends or colleagues would enjoy this newsletter, go ahead and forward this email to them. This will help us spread the word!

Semper discentes,

The D4S Team