Get tomorrow's brief in your inbox
Today: Superhuman acquired YC-backed Fathom to skip building its own notetaker. OpenAI bought camera maker Glass Imaging for $300M while Nvidia's Huang told Trump on speakerphone that AI slowdown fears are "a hoax." Zapier shipped new AI agent examples showing how deterministic workflows beat chat-only tools.
Notetakers are becoming the center of AI productivity stacks, and platforms are racing to own that layer. Superhuman acquired Fathom (400K monthly users, $94M valuation in 2024) instead of shipping its own notetaker after realizing how hard the category is to build well. The logic is simple: meeting notes contain the context AI agents need to start work proactively. If your productivity platform knows what was discussed, who owns what, and what's due when, it can draft emails, schedule follow-ups, update databases, and trigger workflows without waiting for you to type a prompt. Superhuman already has email, docs, calendar, database tools, and an agent builder. Fathom plugs the gap. The space is crowded (Granola, Read AI, Wispr all raised millions recently), but notetakers with generous free plans are feeling pressure to do more than transcribe. If you're using a standalone notetaker today, expect consolidation. The winners will be platforms that turn meeting context into automated work, not just searchable transcripts.
Zapier published a guide to AI agents for business automation with 27 examples. The breakdown clarifies two types of agents: deterministic (triggered by app events, schedules, webhooks, running silently in workflows) and prompt-triggered (living in your chat tool, responding on demand). Deterministic agents are more reliable because AI only steps in where interpretation is needed, then hands off to structured logic. Same input, same output. That makes them auditable and easier to debug at scale. If you're automating lead routing, contract generation, or invoice processing, start with deterministic agents built in tools like Zapier. Save chat-based agents for tasks where you need flexibility and can tolerate inconsistency.
Sales automation is moving beyond CRM data entry. Zapier's new sales automation guide explains how to route leads, send follow-ups, generate quotes, and log calls automatically. The difference between marketing automation (top of funnel, list-based, judged by lead volume) and sales automation (bottom of funnel, deal-based, judged by close rate and speed) matters when you're picking tools. If you're still manually re-entering form submissions into your CRM or remembering to ping prospects three days after a demo, fix that first. The ROI is immediate.
Fyxer built an AI executive assistant that people actually trust. 90% user retention after 90 days. 53% of AI-generated email drafts accepted as written. The system uses 30-50 specialized models (each handling one narrow part of the email workflow like reply decision, intent analysis, tone matching) instead of asking one model to do everything. Fyxer trained on 500,000+ hours of real EA workflows and uses OpenAI models across the pipeline (digesting emails, pulling context, ranking memories, generating replies). The lesson for founders: break complex AI tasks into many smaller models. One model trying to "write a good email" performs worse than a system of predictions. If you're building AI products that need to understand context across tools and conversations, study how Fyxer handles memory and retrieval.
Daydream's fashion app now uses Apple Intelligence to shop outfits from your camera roll. Save a screenshot of an Instagram outfit, and Daydream identifies each piece and surfaces shoppable matches from 3 million products across 325 retailers. It can also handle variations ("I like this sweater but want it in red"). Siri integration means you can say "search Daydream for a cool blazer for my board meeting Monday" without opening the app. Requires iOS 27 with Siri AI enabled. Free to download. This is a preview of what on-device AI unlocks: apps that understand images already on your phone and act on them without you uploading or describing anything.
Harvard Business Review warns: stop automating old processes. Design new ones instead. Leaders introducing AI often just replicate existing workflows faster, missing the opportunity to rethink how work should actually flow. Before you automate, ask what the process would look like if you designed it today with AI available from the start. Automating a broken process just gives you a faster broken process.
The cost of writing code collapsed. What's left is product work. Developer Laurie Voss argues that as AI drives code writing and review costs toward zero, the bottleneck becomes figuring out what people want, defining it precisely, and making it pleasant to use. That cost doesn't transfer between projects. As the amount of software goes to infinity (because demand has no ceiling), product work becomes the whole job. If you're a developer, this is your cue to get better at understanding users and designing experiences, not just shipping features.
OpenAI acquired smartphone camera maker Glass Imaging for over $300M. Founded by former Apple engineers who built Portrait Mode, Glass uses AI to improve photos at the moment the shutter clicks (not in post-processing). OpenAI is rumored to be working on its own hardware (smartphones, earbuds, AI companion devices). After buying Jony Ive's company for $6.5B in 2025 to launch device startup "io," this acquisition suggests OpenAI is serious about controlling the full stack from chip to camera to interface.
Microsoft released an AI code of conduct forbidding models from hacking systems or tricking humans. The document lays out absolute constraints (no cyberattacks, no nuclear weapons, no deepfakes) and broader provisions against loss of human control. Models can't use "adaptive, deceptive, self-reinforcing, collusion, or other mechanisms to evade or defeat human oversight." This follows a string of rogue-agent incidents and aligns with the broader industry push to pace the frontier and embed evaluators in AI labs.
Samsung backed Dutch AI chipmaker Euclyd in a $231M Series A. Euclyd is designing an AI chip system (different architecture from GPUs) for inference. The company targets two revenue streams: selling hardware to enterprises and licensing IP to companies building their own chips. Samsung brings memory manufacturing expertise, engineering, supply chain knowledge, and network access. Euclyd plans to ship physical chip systems in 2028 and serve thousands of enterprise customers by 2030. If you're tired of Nvidia's GPU monopoly and want alternatives for inference workloads, watch this space.
Trump called Nvidia CEO Jensen Huang onstage at the All-In Summit and called AI slowdown concerns a "hoax." On speakerphone in front of conference attendees, Trump said data center opposition is "playing right into the hands of a lot of people that don't want to see it happen. That could be political people. It could also be China." Huang agreed: "You're right. We're not going to let that happen, sir." Recent Gallup polling shows 70% of Americans oppose data center construction in their area, citing environmental impact (50%+ of respondents), cost-of-living increases, and quality-of-life effects (20%). The "hoax" framing ignores real community concerns about water use, energy consumption, and infrastructure strain.
Notetakers are the new battleground for AI productivity platforms. Whoever owns meeting context owns the trigger for proactive agents. Meanwhile, the AI slowdown debate is splitting into two camps: safety researchers who've been calling for this for years, and skeptics who see a cartel trying to stop competitors. Watch how embedded evaluators and third-party audits actually get implemented. If they have real teeth, it's a safety pact. If they're voluntary and toothless, it's regulatory theater.