Jul 27, 2026 · 5 min listen · Last updated July 27, 2026
From storyflo. This is your daily audio brief. Theo here. July 27th, tech desk. Five stories from the last twenty-four hours — here's where I'd start. Let's get into it. First, from BAIR Blog. Teaching LLMs to Update Beliefs for Efficient Long-Horizon Interaction.
Listen · storyflo · A.I.
Daily A.I. Brief · July 27th
0:00-5:26
Pick your daily storyteller
Subscribe to match with Theo, Jessica, Chloe, Mason, Brock — your voice, every brief.
Audio pre-rendered by Storyflo · cached + delivered from the edge
Teaching LLMs to Update Beliefs for Efficient Long-Horizon Interaction
I’ve been digging into this new ABBEL framework and the thing that really sticks out is how they treat the summary as a belief state that gets updated step‑by‑step, instead of just shoving the whole history into the model. By pulling the “what to remember” out of the raw context and supervising the content of each belief, they turn the summarization problem into a clean, isolated task.
The clever part is the belief‑grading signal. They let the model act as both encoder and decoder: after each action they check how well the updated belief can reconstruct the latest observation. That gives a concrete reward for keeping the right pieces of information, and it works even when you can’t hand‑craft a perfect heuristic.
In practice they showed it cuts the memory footprint dramatically while closing about half the performance gap to full‑context models on a collaborative coding benchmark. On the CombinationLock game, adding domain‑specific grading even let ABBEL beat the full‑context baseline.
Overall, the takeaway is that by turning the summary into a supervised belief code, you get a more efficient learning signal and a lot less context to juggle, which seems especially useful for long‑horizon tasks like coding assistance.
America’s AI Investment Boom Is Reshaping the Economy
The surge in AI spending isn’t just about new models—it’s the massive build‑out of data centres, high‑speed fibre, and power that’s pulling in billions for construction, engineering and chip production. That hardware push is spilling over into traditional sectors: factories are using AI to catch defects early, hospitals are shaving admin work, and banks are crunching market data in seconds.
Those upgrades are reshaping how investors read the economy. Currency markets start to price in higher productivity and capital spending long before official GDP numbers appear, so the dollar can move on AI‑related announcements alone.
Because the real value will come from how businesses embed the tools, firms are spending on training, governance and data security to turn short‑term efficiency into lasting productivity gains.
The ripple effect reaches places like Texas, Arizona and Virginia, where reliable power and skilled labor attract the new data‑centre projects, boosting local jobs and feeding back into the broader U.S. economic outlook.
How AI is shortening drug discovery timelines in China
I’ve been thinking about how Insilico’s pipeline actually works under the hood. They stitch together a generative‑AI engine that spits out target proteins and molecule sketches, then hand those designs to researchers who synthesize a handful—often under a hundred—rather than the thousands you’d expect. The lab in Shanghai runs the wet‑lab validation, with automated sampling and screening that keeps the feedback loop tight.
Because the AI narrows the chemistry space, they can name a preclinical candidate in about nine to thirteen months, whereas the classic route drags on for four‑plus years. Their fastest run hit nomination in nine months, and they’ve already logged thirty‑one preclinical leads since 2021, with thirteen IND clearances.
The Chinese infrastructure matters too: lower operating costs, a fast‑track review system, and a big‑scale lab capacity let them shave roughly two years off the usual timeline. That’s why a Pfizer exec says trials there can be three times faster and half the cost of Europe.
On the people side, the company is reshaping its workforce—about forty percent of the software side could be automated or repurposed, so scientists are learning to steer the AI and the robots instead of doing every assay by hand. It’s a quiet shift, but it’s what’s driving the speed.
Building the enterprise environment for agentic AI
I’ve been digging into Intel’s take on building an enterprise environment for agentic AI, and the thing that kept popping up is that it’s not just about a bigger language model—it’s a full‑stack systems challenge. They ran thousands of workload experiments and found that the real bottleneck shows up when you look at how many agents you can squeeze onto each virtual CPU, not just the raw number of agents.
What surprised me was the shift in what you should be watching: average CPU usage looks fine, but agents are bursty, hopping between waiting for model replies and short compute spikes. The sweet spot is to track the 95th‑percentile task latency; that tells you when queues are forming before the average task time even creeps up. And when you need more capacity, the data says “scale out” – add more nodes – is usually the better move, reserving “scale up” for those heavy‑compute, memory‑tight cases.
Intel extended their open‑source Terminal‑Bench harness to separate the agent’s orchestration overhead from the LLM’s variability, replaying deterministic LLM responses. That let them profile a wide mix of tasks—from compilation and database ops to video transcoding and ML training—so the findings feel relevant to real‑world enterprise workloads.
Bottom line: successful agentic AI hinges on planning around agent‑per‑vCPU density, monitoring latency, and scaling out by default, all while wrapping the automation around already codified, measurable workflows like code generation, test farms, ticket triage, and security reviews. It’s less about flashy model metrics and more about a reliable, cost‑aware foundation that lets the agents do the real work.
What’s shifting under the hood is that AI is no longer just a helper for sifting through massive libraries; it’s actually designing molecules from scratch and feeding those designs straight into the wet lab. That flips the old workflow—instead of testing millions of random compounds, researchers get a tighter set of candidates that already look promising, which means the bench teams now have to handle more detailed, data‑rich experiments. The bottleneck has become the quality of the data itself: public datasets are full of positives, missing the failures that would teach models what not to do, and even some images get subtly altered. Vendors are starting to lock down data integrity with hash‑based checks, and the bigger vision is an integrated loop where every assay automatically updates the AI, keeping the whole system FAIR and continuously improving. The net effect is a faster, more informed cycle that could finally bring AI‑designed drugs into the clinic within a few years.
Imagine a healthcare system made up of multiple AI agents: one that manages symptom assessment, another scheduling, a third insurance, and a fourth pharmacy.
Each is an expert in its domain. But they all have their own distinct knowledge and objectives. Today they can exchange data, but they are not yet able to actually coordinate patient care without a human making the decisions.
“The intelligence is already there.
METR introduces a new metric to calculate exactly when AI agents become more expensive than humans
So I was looking into this new metric that's supposed to help figure out when AI agents become more expensive to use than humans. It's called the expenditure horizon and it basically puts a dollar figure on how cost-effective these AI agents are at solving problems. The idea is to give a clear picture of when it's more cost-effective to use AI versus human labor.
The early results from testing this metric on a specific AI model, the NanoGPT speedrun, weren't super impressive. It seems like the metric has some blind spots, which means it might not be giving the full picture. This could be because the metric is still new and needs some refinement, or it could be that the AI models being tested are just not complex enough yet.
One thing that's interesting is that the newest generation of AI models could totally change how this metric works. If these newer models are more efficient or can solve problems in different ways, it could affect how the expenditure horizon is calculated. This means that even if the early results are underwhelming, it's still worth keeping an eye on this metric to see how it develops over time.
It's also worth noting that the expenditure horizon is not just about the upfront cost of using AI agents versus human labor. It's about the overall cost-effectiveness of using these agents to solve problems. This means that even if the initial cost of setting up an AI agent is high, it could still be more cost-effective in the long run if it can solve problems faster or more accurately than humans.
Overall, this new metric is an interesting development in the field of AI research. While it's still early days and there are some kinks to work out, it could potentially give us a better understanding of when and how to use AI agents in different contexts.
How I Reproduced BM25, Dense Retrieval, and SPLADE on a 16GB MacBook
I'm telling you, it's all about the math behind the scenes. So, the author of this post took on the challenge of reproducing three retrieval baselines on a 16GB MacBook - BM25, Dense Retrieval, and SPLADE. What's crazy is that BM25, which is a pretty established model, crashed multiple times due to memory issues. The author had to dig into the code and optimize it to get it working. They also ran into issues with Dense Retrieval, which required a lot of memory and computation power. But the real kicker was SPLADE, which is a more recent model that's supposed to be more efficient. It turned out that the author's MacBook just wasn't powerful enough to handle it, and they had to use a workaround to get it to work.
The Most Beautiful Statistic: The History and the Science of the Humble Mean
You know how we always talk about averages and how they're not always the best way to understand a dataset? Well, it turns out the mean has a pretty rich history that's worth exploring. The ancient Greeks were actually one of the first to use the mean as a way to describe the center of a set of numbers, but it wasn't until the 17th century that the concept really started to take shape. A mathematician named John Graunt used the mean to analyze mortality rates in London, which was a pretty groundbreaking move at the time.
Fast forward to the 19th century, and the mean becomes a crucial tool in statistics. Mathematicians like Carl Friedrich Gauss and Adolphe Quetelet start to develop the concept further, using it to study everything from weather patterns to human height. And it's not just about the math – the mean has real-world implications. For example, it's used in finance to calculate the average return on investment, and in medicine to understand the average lifespan of a patient.
But here's the thing: the mean is not always the best way to understand a dataset. Sometimes, it's skewed by outliers or extreme values, which can give a misleading picture of the data as a whole. That's why we have other measures of central tendency, like the median and the mode. But the mean remains a powerful tool in statistics, and its history is a fascinating story of how math and science intersect.
One of the most interesting things about the mean is how it's used in different fields. In economics, it's used to calculate the cost of living index, while in psychology, it's used to study personality traits. And in sports, it's used to calculate batting averages and other performance metrics. The mean is everywhere, and its versatility is part of what makes it so beautiful.
So the next time you hear someone talking about averages, remember the rich history behind the mean. It's not just a simple math concept – it's a powerful tool that's been shaped by centuries of scientific inquiry and mathematical innovation.
Shared Claude chats were reportedly showing up in search engines
So I was looking into this thing where people were using Anthropic's Claude chatbot and having these private conversations, right? And it turns out that some of those conversations were actually showing up in Google search results, which is pretty surprising. The reason this was happening is that the pages where these conversations were taking place didn't have a thing called a noindex tag, which is like a flag that tells search engines not to index that page. As a result, some pretty sensitive stuff was appearing in search results, like crypto keys and legal questions. It's not the first time this has happened, either - OpenAI had a similar issue last year.