AI + Agents || Science · Economics · Data · Trust · Safety · Builder || @uber · @google · @cisco · @stanford · @ohiostate || Neurospicy || views my own
Sunnyvale, CA
Joined May 2009
- Tweets2.2K
- Following7.5K
- Followers687
- Likes3.3K
Pinned Tweet
Wrapped up #NvidiaGTC feeling inspired and grateful! 🚀 Congrats to the organizers. Amazing seeing #NVIDIA Blackwell and DGX SuperPOD capabilities. From groundbreaking exhibits to enlightening conversations, excited for the future of AI.
#GTC2024 #AIWoodstock – at San Jose, CA
The next big opportunity in AI is access to real enterprise data and workflows in use.
As models get better, generic data becomes less useful, and specialized industry knowledge becomes more valuable.
One thing to keep in mind is that we also need better steering and evals to test that AI against what actually happens in the real world.
a few of my predictions for the next 6 months of ai data:
- a monopolistic human data buy side layer emerges. right now you can use zap mail, heyreach, instantly and claude max plan to reach out to practically every potential human data buyer on the world in 7 days and (even with the worst strategy) at least come out with roughly 20 buyers spanning rl builders, neo labs and (if lucky) one or two major labs, who you could easily offload $20-$50m cumulatively from the very start. it is not going to be this easy to find buyers forever. as labs inevitably get f*cked by bad apples on rights, de-id, QC and the rest, a select few will emerge as “transaction gates” in which every corpora must flow them in order for the labs to deem as safe to buy. analogous to how a few labs already don’t touch corpora if it was de-id by Tonic Textual.
- the waterline continues to rise. ai labs are buying real company data to build artificial workspaces to teach their models how to do an office job. using a headcount as a metric, 3 months ago they bought company internals with 10+ headcount, now they almost exclusively purchase from companies with 50 or more employees. as models saturate these smaller workspaces, demand for smaller company corpora plummets whilst demand for the already dwindling supply of large sprawling company internals increases. yet doesn’t skyrocket as at a certain point if the price becomes too high then labs fall back on synthetic expert work. it goes without saying that small companies in a esoteric domain models underperform in that hold vertical depth (e.g a 30-man freight brokerage) will still hold immense value.
- reality gets graded more and more. as the most knowledgeable domain experts begin to fail in writing valid rubrics that aren’t instantly saturated we are going to have to get reality to play the verifier, domain expert and environment, if we want progress to continue. rubrics shift from the attributes of a perfect NDA to a lab measured critical tempurature.
- atom environments. why hasn’t the vertiginous increase of ai capabilities led to huge economic growth? how can we get non-tech people to understand how intelligent ai is already? how do we find the best value/token ratio? to “solve” these questions I believe need to give ai a “body” in which it can interact with the real world at extremely low bandwidth. giving ai a body sounds like a humanoid which is obviously just one small slice of real world interaction, a little bit like how literally none of the ways we interact with in the real world is only via our real body but via tools instead like handles, vehicles, instruments and cutlery. maybe I should reframe it from giving ai a body to giving the world ai
- protocols for company data emerge. for such a fuzzy yet crucial area of ai, it is insane how unorganized certain steps in company data brokerage are, such as giving context on what your companies internals actually hold.
With so much AI generated slop flooding social media, we should have a better controls for a human only feed?
I get that distinguishing AI generated content from human content is hard. But with AI, shouldn’t we be more ambitious, maybe have it even as a premium feature?
Curious what others are seeing. Would you want this?
@Pinterest @tiktok_us @instagram @Reddit @YouTube @X
Think one can vibe code an entire product over a wknd? Here’s a great example of what it takes to build and maintain a real world product
every week someone tells me they’re going to vibe-code their own @beehiiv alternative over the weekend.
wishing them the best, here’s the backlog of things to make sure to include:
> 99.3% email delivery rate (industry-leading)
> team of deliverability experts with decades of experience
> infrastructure to send 3M+ emails per minute
> validating every single email submission (1st and 3rd party)
> automated bot-detection to prevent spam signups
> registering for every mailbox provider’s feedback program
> ingesting data from said feedback loops and building suppression lists
> tagging every email so gmail reports complaint rates per newsletter
> classifying every bounce (hard, soft, mailbox full, DNS failure) and knowing which ones mean “try again” vs. “never again”
> three layers of suppression: global, per-publication, and transactional
> temporary suppressions that expire on their own, and permanent ones that never do
> one-click unsubscribe, the way gmail and yahoo now require it
> warming new IPs and domains on a schedule
> watching blocklists around the clock, and pulling an IP from rotation the moment it’s listed
> retrying deferred mail for 24+ hours before calling it a bounce
> blocking disposable, invalid, and typo’d addresses before they ever reach the list
> rate limiting every signup endpoint a bot can find
> filtering bot clicks out of every report (something we pay an ungodly amount of money for)
> an acquisition source on every subscriber
> engagement history for every reader, across every post
> multiple sender addresses for different emails
> custom domain configuration that is aligned with DKIM, SPF, DMARC
> a true WYSIWYG editor: what you write is what lands in the inbox, with no HTML required
> turning that editor into table-based, inline-styled email HTML that holds up in outlook
> real-time collaborative editing, so the whole team can work on the same issue at once
> dynamic content: different sections for different readers, all in a single send
> separate email and web versions of every post, from one draft
> free and premium versions of the same post, split by a paywall break
> merge tags that personalize every email, and a simulated subscriber to preview them before you send
> templates with saved colors, fonts, and spacing, plus a default style for every post
> columns, tables, buttons, blockquotes, sections, file attachments, and a table of contents with anchor links
> native embeds for X, youtube, tiktok, instagram, bluesky, stocktwits, and podcast episodes
> polls, referral blocks, recommendations, and ad slots, built right into the post
> an AI writer, AI images, and a built-in AI agent one keyboard shortcut away
> images resized and optimized automatically, however they’re uploaded
> a live email-size gauge that warns you before gmail clips your newsletter
> custom HTML blocks, cleaned so a stray script tag can’t break your post
> automatically hiding anything an inbox can’t handle (video, iframes, audio) in the email version, while keeping it on the web
> links and embeds made clickable even in outlook, which ignores links wrapped around tables
> vertical alignment in columns, because a layout that looks off by 4px looks broken to readers
> converting your newsletter to audio versions for readers who prefer to listen
> random cohort sends to randomly split your list to test different content and conversion
> slash commands: type “/” and insert any block without touching a menu
> markdown and keyboard shortcuts for people who never want to touch a mouse
> pasting from google docs with the formatting intact
> search and replace across the whole post
> comment threads right in the draft, so editors and writers stop living in google docs
> footnotes and citations that work in the inbox and on the web
> code blocks, for the technical newsletters
> an AI writer that rewrites, translates, and fixes spelling and grammar in place
> autosave, so a closed tab never costs you a draft
> subject line A/B testing, built into the send flow
> drag-and-drop images, from your computer, your media library, or a URL
> rich-text image captions, plus image titles and alt text for accessibility
> audio files that play inline
> file attachments, so a PDF can ride along with the issue
> sections with their own backgrounds, borders, and spacing
> blockquotes with custom borders, colors, and styles
> content breaks and dividers that render cleanly in every inbox
> a REST API for subscribers, posts, segments, automations, and more
> a send API to create and send newsletters programmatically
> creating, updating, and deleting subscribers in real time
> reading and writing custom fields on any subscriber
> tagging subscribers from your own app
> enrolling subscribers into automations via API
> pulling opens, clicks, and engagement for every post
> fetching post HTML rendered for free vs. premium, email vs. web
> creating posts from a template, then appending or prepending content
> fetching segments and every subscriber in them
> managing tiers and premium subscriptions
> pulling referral program data and milestones
> reading poll and survey responses
> pagination that holds up across millions of records
> rate limits that protect the platform without blocking real integrations
> API keys you can create and revoke anytime
> an MCP server so claude, chatgpt, and other AI tools can work inside beehiiv
> drafting and editing posts from an AI chat
> asking for open rates, CTR, and revenue in plain english
> comparing performance across dozens of posts in one request
> looking up any subscriber's full history and engagement
> building segments by describing them
> creating and editing automations
> generating AI images straight into your media library
> managing website pages, redirects, and SEO
> reading and replying to community posts and DMs
> reviewing ad offers and ad revenue reports
> managing podcast shows and episodes
> creating digital products, coupons, and promo codes
> running subscriber growth and earnings reports
> setting up subject line A/B tests
> deliverability by mailbox provider for any post
> searching docs without leaving the chat
> inviting teammates by email
> admin, member, and contributor roles
> admins who control billing, exports, integrations, payments, and domains
> members with full access minus admin permissions
> contributors limited to drafting posts and automations
> roles at the workspace or publication level
> multiple publications under one workspace
> workspace grouping and search for people on multiple teams
> multiple authors per post, with bylines
> guest authors for one-off contributors
> author profiles with bios, avatars, and social links
> community roles mapped to workspace roles (admin, manager, moderator)
> real-time collaboration and comment threads on every draft
> removing access the moment someone leaves
> publication transfers between owners, with written consent
> secure, single-use magic links to log in from desktop
> a no-code website builder on your own custom domain
> an AI website builder that works from plain-english prompts or screenshots
> AI custom elements for live widgets
> drag-and-drop editing, plus HTML access for full control
> site-wide theme tokens for fonts, colors, borders, and shadows
> global navbar, footer, and signup and login modals
> saving sections, blocks, and pages as reusable templates
> version history with one-click restore
> custom pages with their own SEO settings and paywalls
> redirects, so old links never break
> custom robots.txt
> auto-generated meta titles and descriptions for every post
> google news schema
> pages up to 80% faster on largest contentful paint
> image carousels and grids
> youtube blocks that update when you post new videos
> an automatic page for every podcast episode
> pop-ups and gated content to turn visitors into subscribers
> a visual automation builder, no code required
> triggers based on signup source, behavior, purchases, and engagement
> welcome sequences that start the second someone subscribes
> delays and wait steps
> branching logic based on subscriber attributes
> reusable condition sets across automations
> email steps built in the same editor as your newsletter
> webhook steps that call your own systems, with automatic retries
> enrolling subscribers via API
> per-subscriber journey tracking
> opens and clicks for every email in every flow
> editing a draft version while the live automation keeps running
> comparing draft vs. live before you publish changes
> reordering steps in bulk
> segment conditions based on whether an automation started
> product delivery and follow-ups triggered by purchases
> selling files, links, guides, templates, and ebooks
> memberships and time-based products
> 1:1 coaching sessions and appointments
> 0% platform fees on every sale
> a storefront on your beehiiv website
> products sold inside emails and automations
> free products that double as lead magnets
> product variants
> discount codes and promotion codes
> product reviews, with moderation
> product tags for organizing your catalog
> orders dashboard with buyer details
> sales totals per product
> duplicating a product in one click
> automatic delivery the moment someone buys
> payouts through stripe
> live webinars with up to 10,000 attendees
> ticket sales in 10 currencies
> native streaming with video, screen sharing, and chat
> no zoom required
> up to 5 automated reminder emails
> calendar invites sent automatically
> a recording of every session
> selling the replay as its own product
> 0% platform fees on tickets
> free webinars that grow your list
> discount codes on tickets
> 1:1 coaching bookings
> buyers linked to their subscriber profile
> promoting events in your newsletter and on your site
> orders and revenue reporting for every event
> a community built into your newsletter: channels, posts, and comments
> channels that are free or gated to paid tiers
> syncing newsletter posts into the community
> direct messages between you and members
> threaded comments
> pinned posts
> a moderation queue for reported posts and comments
> banning and unbanning members
> admin, manager, and moderator roles
> inviting subscribers, segments, or specific emails
> surfacing your most engaged subscribers to invite first
> community metrics: new members, posts, and activity
> notification emails with your own from name
> channel ordering and a default channel
> luma calendar integration for events
> discord integration that syncs roles with subscription tiers
> native podcast hosting with zero platform fees
> distribution to apple podcasts, spotify, and youtube
> an automatic webpage for every episode
> automatic transcripts
> IAB-standard analytics by country, app, and device
> bot filtering on downloads
> downloads and unique listeners per episode and per show
> private feeds for paying subscribers
> bundling your podcast with your paid subscription
> importing from other hosts with episode GUIDs preserved
> RSS redirects so listeners follow you over
> scheduled episodes
> apple podcasts categories and metadata
> podcast players embedded in your posts
> a full episode library on your site
> webhooks for episode and private-feed events
> segments built on behavior, attributes, and engagement
> dynamic segments that update automatically
> static segments that stay fixed
> filtering by acquisition source, including magic links
> filtering by custom fields and tags
> filtering by subscription tier and paid trial status
> filtering by whether an automation started
> reusable condition sets
> recalculating a segment on demand
> sending to a segment, or excluding one
> segments that power dynamic content inside a single post
> segments as automation conditions
> adding a whole segment to a newsletter list
> multiple newsletter lists in one publication
> exporting any segment
> webhooks for new, confirmed, and deleted subscribers
> upgrade, downgrade, pause, and resume events
> tier created, deleted, paused, and resumed events
> newsletter list subscribe and unsubscribe events
> post sent, scheduled, and updated events
> survey response events
> podcast episode published, updated, archived, and deleted events
> private podcast feed access events
> choosing exactly which events each endpoint receives
> a signing secret on every endpoint
> automatic retries on failed deliveries
> a log of every delivery attempt
> manually retrying any failed message
> disabling an endpoint without deleting it
> versioned payloads
> webhook steps inside automations
> survey forms with multiple question types
> polls that readers answer with one click, right in the inbox
> every answer tied to a real subscriber
> live poll results
> per-question response breakdowns
> aggregate results across every survey
> individual responses you can browse one by one
> webhooks the moment a survey is submitted
> survey and poll data available via API
> polls on the web and in email
> signup flows that send new subscribers through a survey
> surveys that tell you what to write next
> polls for picking topics, products, and pricing
> response data you can turn into segments
> no third-party form tool required
> a library of professionally designed post templates
> custom colors, fonts, and spacing on every template
> a default style applied to every new post
> content outlines baked into templates
> separate email and web styling
> theme design tokens per template
> duplicating templates in one click
> a custom email footer applied to every post
> templates with sponsor slots, referral blocks, and recommendations built in
> creating posts from templates via API
> template changes picked up instantly by API-created posts
> website templates
> saved sections, blocks, and pages for your site
> welcome email templates
> templates your whole team shares
> the largest native ad network for newsletters
> brands like netflix, hubspot, notion, and nike (and 525+ others)
> ads available to claim for your upcoming posts
> dropping an ad into a post with one click
> earnings and payouts reported for every placement
> ad revenue reports post by post
> a history of every opportunity you accepted or declined
> bot-filtered clicks advertisers can trust
> custom link parameters for advertiser attribution
> brand safety review on every campaign
> payouts to your beehiiv wallet
> direct sponsorships: selling your own placements
> sponsorship products with your own formats and pricing
> scheduled sponsor placements with advertiser details
> 0% platform fees on paid subscriptions
> multiple paid tiers
> monthly and annual plans
> free trials
> paid trials, like $1 for 30 days
> metered paywalls, new york times-style
> a paywall break inside any post
> discount offers
> gift subscriptions
> group subscriptions: one buyer, many seats
> custom group deals for teams and companies
> pausing and resuming subscriptions
> private podcast feeds and gated community channels as paid perks
> payouts through stripe connect
> recommendation network with 60K+ active newsletters
> getting recommended by thousands of other newsletters
> free, reciprocal recommendations
> paid recommendations: pay to grow, or get paid to recommend
> discovering newsletters to partner with
> AI-suggested recommendation partners
> applications for paid recommendations, with accept and reject
> engagement and spend insights for every partner
> pausing, resuming, or ending any partnership
> fraud detection on paid recommendations
> clawing back payouts for fake subscribers
> recommendations inside your newsletter, not just at signup
> attribution for every subscriber a partner sends you
> one-click subscribes powered by magic links
> one-click subscribe links, no form needed
> UTM parameters carried through on every link
> rate limiting on every request
> bot detection and honeypots
> a confirmation step only when a click looks suspicious
> double opt-in respected
> redirecting new subscribers straight to your site
> magic links as a filterable acquisition source
> links you can drop into other newsletters, socials, and ads
> login magic links so paid subscribers can read premium posts
> single-use tokens that expire in minutes
> rate limits on login requests
> protection against timing attacks
> logging in on desktop from a link sent to mobile
> faster logins across every beehiiv publication
> embeddable subscribe forms for any website
> full-width, fixed-width, and slim embeds
> an attribution script to track where signups come from
> custom colors, fonts, and styling
> theme design tokens per form
> a draft and publish workflow for form changes
> comparing draft vs. live before publishing
> multiple forms per publication
> custom success behavior: a message or a redirect
> pop-ups on your beehiiv site
> gating posts to subscribers only
> readers picking which newsletter lists to join
> signup flows that run right after subscribing
> bot protection on every submission
> forms that work on wordpress, webflow, and anywhere else
> a customizable link-in-bio page
> hosted on your own beehiiv site and domain
> a subscribe form built in
> your brand colors, fonts, and buttons
> social links for every platform
> links to your latest posts
> links to your products and paid tiers
> a linktree integration that captures newsletter signups
> every visitor a potential subscriber
> traffic from instagram, tiktok, and X turned into owned email subscribers
> visits tracked in your website analytics
> signups tagged with their acquisition source
> mobile-first design
> no third-party link tool required
> updated in seconds, from the same dashboard as everything else
> real-time website analytics, no third-party tools
> traffic by channel
> filters for location, device, OS, and referral source
> website and newsletter analytics in one place
> breakdowns by page, referrer, and more
> custom conditions to slice any report
> AI crawler and bot analytics: what was served vs. blocked
> control over which crawlers can index your site
> UTM tracking on every link
> custom link parameters at the publication level
> an acquisition source on every subscriber
> google analytics integration
> podcast analytics alongside your site
> trends over time
> website traffic connected to the subscribers it produced
> native integrations with linktree, hubspot, manychat, meta, and dozens of other apps
> human support team with a dozen experts ready to assist with whatever you need
oh, and besides having to manage all of this on your own indefinitely forever… your vibe-coded solution will still require paying for and stitching together a ton of software:
> email sending (sendgrid, amazon SES, postmark)
> web hosting (vercel, AWS, render)
> database + queue (supabase, neon, RDS)
> payments (stripe, memberful)
> auth (clerk, auth0, supabase auth)
> monitoring + alerting (sentry, datadog, pagerduty)
> workflow engine (inngest, temporal, upstash qstash)
> bot protection (cloudflare)
> email validation (zerobounce, kickbox)
> email rendering (litmus, email on acid)
> click verification (omnivery)
> media storage + CDN (S3, R2, vercel blob)
> product analytics (posthog, mixpanel)
…but at least you saved $59/mo not paying for @beehiiv anymore 🫡
love the idea of a ‘Made by Me’ app section for personal use!l esp.
@AppStore @GooglePlay if this is in the works 👀
Imagine a system that writes its own code, runs its own tests, and approves its own deployment. What could go wrong?
Satya raises a big question about trust as agents get smarter.
These models can reason, write code, access sensitive data, and take real actions. But who reviews what’s allowed? Who checks if they got it right?
I think the next big engineering challenge is the control layer around them. Permissions, architecture rules, independent verification, and the ability to stop things *before* they go sideways.
Especially with coding agents. We’re getting really good at generating code, but who’s independently checking the architecture, dependencies, and what’s breaking downstream?
The big shift happening in software is making the UI optional.
Headless platforms let AI agents discover, reason, execute, transact, and pay for software directly.
As agents are becoming customers, we need to be designing software ecosystem for them.
Core ideas around always on agents
How do always-on, proactive agents like @Bot work?
Watch my lecture at Stanford CS146S:
2:35 How we got here
6:13 What changed in the models
10:06 Inside an always-on agent
18:07 The harness
27:26 Context engineering
34:26 Where this is going
The next evolution of AI is not simply better general intelligence. It's specialized intelligence with that can deliver reliable domain outcomes
underrated resources for tech adjacent roles
I see 98% as creating opportunity for everyone.
We’re still in the early innings. As AI becomes more useful in everyday life, the market will only get bigger. So much room for innovation and experimentation.
98% of US households aren't paying for AI yet
More charts in State of Markets II: a16z.news/p/state-of-markets…
Same reason I’m trying all of them right now to understand what they can do, but ultimately betting on open source ecosystem.
Models will keep getting cheaper and better. The bigger question is who ends up owning the data, memory, and context. That’s where the real value will accrue.
there’s something quite awkward about all the personal agents in the current hype cycle
muse, grok bot, instinct, dots, and whatever google, anthropic will come up with
none of them is “mine”
i’d be trusting a vendor for some of the most sensitive data about me and bet on them handling it with extreme responsibility. what if they have a data leakage incident? what if they hired a wrong employee? what if they simply break bad?
i’d also be betting on their model. right now i’ve set up a lot of my stuff in grok bot, but what if their model falls behind? what if another model becomes 10x better on pure intelligence? then i’ll either miss out on a better assistant, or have to bite the bullet and do a full migration
i’d also have to bet on their product. they may not build the features i want. they may not allow me to customize enough. and one day my assistant may show me ads
that feels like way too much betting than what i’d be comfortable with, if i truly want literally everything in my life to go through and be taken care of by an assistant
with all that considered, i’m more bullish on open source personal agents that run on people’s own computers. but i think these open source projects have to break out of the assumption that their users are developers and they can just throw a repo at them and ask them to launch a terminal
a well polished, fool proof, community maintained, vendor agnostic, free personal agent that everyone can run by themselves without locking into a SaaS - that’s what i think many people will need
AI is making it easy to create knowledge.
The real advantage comes from how quickly we can understand it, question it, and apply it when it’s needed.
We'll be spending a lot more time trying to understand the outputs of language models. A few thoughts, tips & tricks:
Writing. Something I've had success with: Ask your LLM to explain something in ASD-STE100, it's a controlled language specification originally developed for aerospace maintenance documentation. LLMs well-versed in this language and it comes with heavy constraints on clean writing style that I often find a lot more readable. Sometimes I've tried to soften it a bit e.g. ask for "80% of the way to ASD-STE100" because the spec is quite stringent. But even better:
Diagrams / images. Instead of writing, ask your LLM to create a diagram. These can be a lot easier to process, parse, and understand. But even better:
Web pages. Ask for output "in HTML" to get a beautiful, interactive webpage. LLMs are getting really good at frontend and can create beautiful experiences, animations, etc. But even better:
Explainer videos. The output format I am most bullish on is fully custom / bespoke explainer videos generated on any arbitrary topic. Experiment with things like "Create a 3b1b style video explainer on X. Use my ElevenLabs API key for audio narration". (you'd need an API key for the latter or you can ask your LLM to find you decent free alternatives that use your local compute). This is actually starting to work!
In summary:
- As LLMs get better, they will do more and more of the legwork autonomously, and a lot more of our work will rise up the abstractions into oversight and understanding.
- Luckily, LLMs can help here too because as intelligence and code are increasingly abundant, you can ask for large, custom, discardable software artifacts (e.g. web apps, video explainers) that would have never made sense to create before. Push the boundaries here and you'll be surprised.
One idea stuck with me yest at The AI Conference.
We spend so much time trying to make individual models trustworthy. But why should trust depend on one model getting everything right?
Different models have different strengths and blind spots. Let them challenge each other.
But three models agreeing also doesn’t mean they’re right. They can all be wrong together.
The real challenge isn’t getting models to agree. It’s independently verifying what they say and do.
Trust needs to be built around models, not just inside them.
@AIconference