You.com

We're a week away from Step SF—and this year it's bigger than ever. 📅 August 27, 2026 | 📍 The Midway, San Francisco Founders, investors, builders, and operators are coming together for a full day of unfiltered conversations, live product demos, and real networking. What's on deck: 🎤 Main stage talks from founders and investors shaping AI and tech 🖥️ An expo floor packed with the next generation of intelligent tools 🤝 Pre-matched 1-on-1s between founders and VCs 🚀 A live pitch stage in front of top-tier investors 🎯 Mentorship roundtables with operators who've actually scaled Plus satellite events all week—including a pitch competition hosted by You.com and Amazon Web Services (AWS) on August 26. If you're building or backing what's next in AI, this is where you want to be.

A bigger context window looks like pure upside on a spec sheet. It's actually two problems hiding behind one number: a cost curve and an accuracy cliff. The cost side: attention scales quadratically. Go from 4K to 128K tokens—a 32x increase—and compute per layer jumps roughly 1,024x. The accuracy side is worse. NVIDIA's RULER benchmark found only half of 17 models held up at 32K tokens, despite all claiming 32K+ support. GPT-4 lost 15.4 accuracy points between 4K and 128K. Gemini 1.5 Pro lost just 2.3. Bigger windows aren't the fix. Context budgeting is. Determine what earns space in context, what gets trimmed, and whether retrieval beats stuffing the whole corpus in. Position matters too—highest-relevance content should load first, not get buried in the middle. Learn more on the mechanics, current model tiers, and how to budget tokens instead of filling the window 👇

Agents are only as good as what they can actually see. Alex Triplett, our COO, is digging into this gap on stage at AI Summit Seoul. Agentic AI deployments often fail because the agent is reasoning over stale or incomplete information. You can chain all the tool calls you want, but if the retrieval layer underneath is blind, the output is a confident guess wearing a trench coat. 🤡 📍 Catch the talk: "The Blind Spot of Agentic AI: Why Your Agents Are Only as Good as What They Can See" 🗓️ Aug 19, 1:30–2:00 PM We'll also be at Booth F400 all show—come by if you want to talk through what's breaking in your agent stack. See you in Seoul. 🇰🇷

In two weeks Saahil J., CTO of You.com, will be taking the stage at API World! Saahil's talk will dig into a question that matters more every day: as AI agents become the fastest-growing consumers of APIs, is our infrastructure ready for them? API systems were originally built for human developers, not autonomous agents making decisions and calls at scale. He'll unpack: 🔓 Why the API economy is at risk of collapsing into closed silos as major platforms wall off access ⚙️ What agentic API infrastructure actually requires — from endpoint design to traffic management 📈 The business case for staying open in an agent-driven world 🏗️ How developers and companies can help build the open foundation layer that prevents platform lock-in Saahil brings a rare mix of research depth and applied experience to this conversation. If you're thinking about how the API economy evolves in the agentic era, this is a session you won't want to miss. 📍 Santa Clara, CA, September 2, 4PM PT catch Saahil live at API World—more details below.

Every LLM has an expiration date—the day its training data was frozen. 🕰️ Ask it about a regulation that got repealed last month, or a price that changed last week, and it'll answer confidently anyway. That's the absence of a system. A web search API is the fix. It lets your AI query the live web at inference time instead of guessing from memory. Structured JSON back, not a page rendered for humans—titles, snippets, source URLs, timestamps. What separates a good one from demo-ware? How fast it indexes new content, whether every result traces to a real source (no citation, no way to verify—that's a hallucination risk you're shipping to production), and what happens when it fails under load. Read more 👇

Majority of AI search demos look impressive...until they're asked to show their work. That's the difference between "AI bolted onto search" and something you'd put in front of a client. Legacy search hands you links, most AI search hands you an answer. But, if it can't say where the info came from, it's a confident guess, not an upgrade. What actually matters is whether or not the system can formulate its own research questions, judge which sources are credible, and cite the exact claim back to its source. We wrote up what's changing—and why citation, not cleverness, is the part worth scrutinizing. 👇

Step SF is back for its third year, and this one's shaping up to be the real deal. 📍 August 27, The Midway, Dogpatch. Join us in two weeks for a full day with founders, investors, and builders who are shipping. What's different about Step? There's no rigid agenda and no formal networking rooms. Instead it's comprised of pre-matched 1-on-1s between founders and VCs, demo kiosks where you get hands-on with the product, mentorship roundtables, and a pitch stage where selected startups get in front of top-tier investors live. We're hosting a pitch competition the night before and there will be satellite events at Frontier Tower and the Tesla Showroom leading up to the big day. Tickets and startup applications are live now. 👇

When evaluating citation-backed answer synthesis, teams focus on the model. But the serving layer shapes your latency, cost, and ability to actually hit your SLAs under load. This is why we built the Answer API on the Baseten inference platform. Open-weight models on managed inference let us deliver predictable latency during traffic spikes and the ability to fine-tune without rebuilding your serving stack at one-third the cost. The numbers: 93.48% accuracy, 2.67s P50 latency.  If you're thinking through how to ship grounded answers at scale, the infrastructure choice matters as much as the model. This post covers the pipeline, the calculation that led us to open weights, handling model selection as the frontier moves, and how Baseten addresses these needs. 👇

Doors open TOMORROW night at the You.com AI Forum in San Francisco, hosted at Databricks. If you've been meaning to request a spot and haven't yet, now's the time. Expect talks from people running production systems, a panel that argues, on record, about where those systems actually break, and a room full of founders, engineers, and researchers who'll push back on your assumptions instead of nodding along. Request to join before the list closes tonight. 🛠️

"Protocol war" is the laziest framing in the agent space right now. A2A and MCP aren't fighting for the same job. MCP connects an agent to its own tools—database, calendar, search index. A2A connects one agent to another agent it doesn't control. By 2026, A2A hit v1.0 with signed Agent Cards, 150+ supporting orgs, and production use at Microsoft, AWS, Salesforce, SAP, and ServiceNow. IBM folded its competing ACP effort into A2A rather than keep a separate standard alive. Agent failures in 2025 had nothing to do with missing interoperability. Weak prompts, sloppy permissions, no retry logic. Wrapping a shaky single-agent system in A2A just adds a network hop and a new place for things to break. A2A earns its place the moment your agent has to reach one your team doesn't control. If everything lives in one codebase with shared memory and shared auth, that's an architecture problem, not an interoperability one. 🧩 Full breakdown from Staff Product Manager, Brian Sparker, below. 👇