35,595 results · Most relevant
The talk's evaluation framework rests on four questions: where can we reduce attention work, where can we store and move less data, how does the serving setup affect user wait time, and how much reasoning the model does. The talk covers how to reduce model computation and how to test the value of those model changes.
numeric details flagged as possible errors in source
Exposing Vanguard's legacy code base to AI tooling via an MCP server was the turning point of the modernization journey, enabling faster feedback loops. Making the legacy code base AI queryable cut the research cycle time from seven days to 15 minutes. Vanguard's product owners, technical leads, and developers all credit the MCP-based AI tool with helping the modernization succeed.
Chris Dreyer picked legal for status because it's a respected profession, then went from legal in general to personal injury based on data showing 70-80% of revenue came from that space — a natural push in that direction.
FleetKit takes a goal, breaks it down, dispatches leaders to do all the research and decide about the approach, puts everything into a human-in-the-loop queue (e.g., GitHub issues that humans can comment on and approve, or that other agents can add context to), and when everything looks tight from an approval standpoint, dispatches builders to do the rest. The ultimate goal is to produce a 'material change' — a change request into their systems or a code change into their repositories.
Frame enterprise AI adoption as a sequence of milestones in a race and use this ladder to diagnose where a company actually is rather than assuming uniform progress. Benchmarks from ~100 enterprises: ~90% have rolled out a general-purpose LLM (Claude or Codex) to employees; ~60% have done training and enablement; only ~50% have rolled out coding agents to the engineering org; and only ~5% have agents in production for non-engineering work. Don't assume an enterprise has coding agents just because they've rolled out LLMs — roughly every other enterprise either hasn't rolled out a coding agent or what they use is closer to not having one.
Exa recently released ExaSnapshot with 400 billion plus snapshots of the web that can be searched by date. Exa is constantly searching the web for the latest information to add to its index, checking for freshness every time. Because it checks URLs for freshness, it can capture a version of those URLs at the point of the check, creating a versioned history. Versioned history can be accessed by providing a URL and a date, requesting the page as it was at or before that date. ExaSnapshot is particularly interesting for the financial services industry.
Exa offers a version-controlled search feature that lets users access the web as it existed at a prior date. Version search lets a user execute a search as if it had been run at a prior point, returning only results that would have been served at or before that point.
The creator explains that TikTok is the only platform where going live exposes you to a bunch of new people who don't know who you are, because the FYP algorithm surfaces your live stream to a random audience. He argues this random-audience effect is a big deal and is why TikTok is dominant.
People who tried AI two years ago ossified that version of reality; if they try it again now, the results are dramatically different. The speaker has run this experiment about 10 times. A Harvard professor dismissed AI based on trying it a few years ago until the speaker had ChatGPT rewrite her 20-year-old dissertation with everything current, and she was amazed. Silicon Valley judges AI by where it could go (the slope of the curve), while most people judge it by what it is today — which is why SV sometimes oversells reality but also sees the trajectory others miss. In the last two and a half months, computer-use and browser-use assistants can do things that were near-impossible six months ago, with success rates jumping from roughly 30% to roughly 90% of the time.
speaker-reported figures; success-rate numbers flagged as possible numeric errors
A restaurant owner can prompt their way to a monthly wholesaler price-change analysis in about five minutes; it costs roughly a dollar a month to run, and even if it finds nothing in 11 out of 12 months, in the month it does find something it can uncover hundreds if not thousands of dollars in savings.
cost and savings figures are the speaker's rough estimates ('like a dollar a month', 'hundreds if not thousands'); five-minute setup and dollar cost carry possible numeric error flags
Under the new agentic-payment open standards, an AI agent must be registered with Visa, Mastercard, or Amex before it can make purchases on a user's behalf. The user gives the agent explicit context — what is being bought and at what amount — so the network can validate that the instruction matches the transaction. Alex Rampell gives a real example: instructing an agent to find a hotel in San Diego and not spend more than $500.
People will have more agents and there will be more agentic things happening, so agents are going to interact with your software in some way, shape, or form. Amazon blocked the Facebook bot (Muse) from ordering things on Amazon — an example of the lock-out approach. Offering an MCP for agents can be a really good solution for a lot of things, because AI doesn't always think it should build everything from scratch; it might decide the best solution is to use a third-party thing that gets what is needed.
Lincoln Murr describes three core agentic trading use cases that exist today: (1) fully managed portfolios where the user specifies an allocation (e.g. 80% S&P, 5% Bitcoin, 15% T-bills) with guardrails and lets the agent run; (2) autonomous execution of scenario-based trades predefined in advance; and (3) translating a high-level natural-language investment idea into actual financial primitives and a strategy.
speaker's framing of current use cases
Al Chen describes that the deployment timeline gets pushed out to a few months when there are formal security reviews, a custom LLM gateway, or RBAC configuration being built, and even further when there is no approval process for pulling images to the registry or no story around sourcing and procuring GPUs.
A-cert is a new Microsoft Research tool, released about a week before the talk, that helps you write judges based off of a spec. You write a spec describing the agent you're building, roughly what you want it to do, your idea of how it should work, and what kind of users it should support; a-cert then generates lots of test cases and fleshes out your spec into actual judging rubrics based on research.
Sam McKenna warns that even after getting to power and approval, a three-month vendor review with procurement, legal, and security can slow down a deal, so sellers should get creative about starting vendor review early in the deal cycle.
P&L transformation after acquiring a company follows a J-curve: start with investment in AI deployment, build the platform, capture workflows, and commit significant team time. Deploying AI requires partnering very closely with the acquired team, iterating, validating the product, potentially training new models, and examining the unique workflows and data within the business. Once past the initial hill, people actually use the product and you unlock a lot of capacity — time savings, usage going through the roof, same output with fewer hours. But capacity and time savings are not the end: you have to put that capacity to work to generate real business impact, which involves incentive alignment, working closely with management teams and trusted partners, helping with growth, and improving N
Doctors and nurses copy-paste between patients and visits; commingled records between patients (e.g., Bob Joe and Bob Joe Jr.) occur fairly often and violate HIPAA; duplicate records in claims data can be 50% or higher depending on jurisdiction; decisions can be made on other people's mixed-in records, especially on high-complexity claims.
moderate
The speaker proposes solving claims problems in three phases: clean data, within-document analysis, and cross-claim factors.
With some amount of certainty from within-claim extraction, you can flag conflicting diagnoses, deduplicate, and search for missing records.
Choose sources → to get a library of your own and a daily brief.