←All presents

The Wake: May 8, 2026

A daily briefing from George's X bookmarks and likes, with source links and older-memory echoes.

The Wake is a daily briefing from George's saved internet. The issue is written as a newsletter first. The tweets are the source material, preserved below for receipts.

Source window: May 7, 2026. Signals: 37 bookmarks and 9 likes.

Brief

We crossed a threshold this week. The debate is no longer whether large language models are helpful. The debate is about how systems built on them will run businesses, browsers, and customer experiences without humans poking the chat box. Anthropic, OpenAI, a dozen tooling startups, and a swarm of developer projects just moved agentic capabilities from R&D curiosities into product-grade infrastructure. Expect a race where the winners are not the best chatbots but the teams that rearchitect workflows, security, and go to market around autonomous agents.

The agent stack just got practical

Two concurrent shifts are visible. First, models are becoming interfaces. OpenAI pushed GPT-Realtime-2 as a voice agent with GPT-5-class reasoning at conversational pace. That turns voice from a novelty feature into a real-time collaborator that can listen, reason, and act. Anthropic released Natural Language Autoencoders, a technique to translate model activations into human-readable text. That gives engineers a way to inspect internal thoughts, detect reward hacking, and build evaluation metrics that are more granular than pass/fail. Those advances are enabling agents that not only respond but explain, evaluate, and self-correct.

Second, the plumbing around agents is maturing. Codex shipped a Chrome plugin and subagents capable of operating multiple tabs, running DevTools in parallel, and keeping results organized. Entire and Printing Press introduced agent-native CLIs and review tooling that make agents reliable parts of CI and developer workflows. Parallel Monitor announced a push-based Monitor API so the web can proactively notify background agents instead of agents constantly polling. Taken together, these are the primitives teams need to build autonomous, networked agents that run real functions.

From chat to function: replacing teams, not tasks

The language in many of these releases and workshops is revealing. Anthropic’s materials are framed less as prompt therapy and more as architecture for autonomous business functions: research, content, customer comms, analytics, and ops that trigger and hand off without human initiation. That is exactly the claim critics feared and the builders wanted. The practical effect is not jobless factories but compression of operational overhead: small teams orchestrating agent fleets that handle the work previously done by many people.

This is why the conversation has moved from how to prompt to how to design agent workflows, checkpoints, review skills, and safety gates. Tools that offer structured outputs, session logs, and review skills are rising because agents require audit trails and programmatic governance if they are to replace whole roles. There is also a growing consensus that model explainability matters. If you can translate activations into text you can quantify unwanted behavior and instrument fixes faster.

Product, market, and the new scarcity

Assume the baseline: everyone will have similar agent tech. Aaron Levie’s point is the right lens. If AI makes the mechanics of building software and routine customer interactions trivial, competitive advantage will shift into areas where humans still matter: customer relationships, deep domain expertise, distribution, and product experience. That changes resource allocation. Expect spending to tilt from pure engineering hires toward sales, growth, and customer success roles that squeeze more value out of agent-driven products.

Founders should also take the founders’ heuristics seriously. Speed and origin stories still matter. Teams that iterate aggressively, respond fast, and treat slop as a deliberate experiment will move ahead. Mitchell Hashimoto’s defense of "slop" highlights the new craft: generate fast, test, and then clean up. But transparency matters. Ship sloppy agent prototypes to internal testers, not to unsuspecting customers.

Security, privacy, and the UX trapdoors

New agent capabilities create new attack surfaces. Browsers and tooling are particularly sensitive. Codex’s Chrome plugin, Dia’s Auto Tools that read cookies and private repos, and Parallel’s Monitor imply agents will increasingly operate with elevated privileges across sites and apps. That convenience accelerates product innovation but also concentrates risk. Watch for cookie-based access models and background monitoring to collide with consent norms and enterprise security controls.

There is also a policy angle. Anthropic’s explainability work is a forward-looking mitigation against reward hacking. But governance is a systems problem, not a model problem. Agent audit trails, review checkpoints, and structured CLI outputs will be the operational controls that enterprises demand before they let agents touch billing, contracts, or personnel actions.

How teams should start

This is not the time for gentle experiments. The opportunity and the risk are both about composition. A few practical moves:

  • Build one autonomous workflow end to end. Replace a whole function in a low-risk area and instrument everything. Measure throughput and failure modes, not just prompt accuracy.
  • Use agent-native CLIs and structured outputs so you can integrate agents into CI and logs. Printing Press and Entire show the way here.
  • Require review skills and checkpoints for any agent that takes external actions. Automate runbooks for human takeover.
  • Treat explainability as a product requirement. If you cannot translate an agent decision into a sequence of evidence and activations, do not let it execute irreversible actions.
  • Shift hiring and GTM resources toward customer engagement and retention. If everyone has equivalent agent capabilities, your defensibility will be in how you own attention and outcomes.

What to watch

  • OpenAI’s GPT-Realtime-2 adoption curve. Voice agents that reason in real time will accelerate agent UX transitions.
  • Anthropic’s NLA tooling and workshop uptake. If the workshop becomes the de facto course for architecting agents, expect a wave of startups rebuilt around multi-agent workflows.
  • Parallel Monitor and push-based web monitors. A move from pull to push for web events changes how background agents are notified and scale.
  • Agent-native CLIs and plugin ecosystems. Printing Press, Entire, and the new CLI projects will define integration patterns.
  • Security incidents tied to browser agents or cookie-based access models. One major breach would set back enterprise adoption.
  • Regulation or enterprise controls around agent auditability and decision logs. Look for procurement demands that require review checkpoints and explainability features.

The change this week is not incremental. Teams that learn to stitch agents into composable, auditable workflows will run far more output per human hour. Teams that treat agents as fancy assistants and leave architecture, monitoring, and governance undone will be overtaken or tripped by their own systems. Build the plumbing, not just the prompts.

Source tweets

Adam.GPT / @TheRealAdamG

Adam.GPT / @TheRealAdamG

  • bookmark: open on X
  • Codex subagents creating and operating multiple tabs in your chrome browser simultaneously.

Malte Ubl / @cramforce

  • bookmark: open on X
  • [Vercel Plugin in Claude/Codex/Cursor] Find the highest active CPU routes in my project and then implement optimization opportunities to reduce usage

cpt n3mo / @cptn3mox

  • bookmark: open on X
  • my normie friends know i work in buy-side and often ask "what're you using codex, claude for?" -> implying that they're still in the LLM era and these agentic tools are just for developers that's a huge misunderstanding because gone are the chatbot days (static action of: user -> interface); as personal agents become the product (dynamic two-way action of: user -> agent -> user) and all we do is talk to a high-context assistant who knows us well enough to handle vague asks, execute online tasks, and eventually act proactively

Aaron Levie / @levie

  • bookmark: open on X
  • When AI makes one thing easy to do, it’s always good to assume that will equally be the case for everyone else. If it’s the case for everyone else, then that means competitive forces will ensure that resources move to new or other areas that create differentiation. If AI makes building software easier, then there will be a relative increase in resources going into sales, marketing, and customer success, because standing out or going deeper with customers else becomes even more important. This will also apply to lots of other areas of work. If you automate getting financial advice and insights, then the differentiation is in client engagement. And on and on. Just ask yourself: if everyone else does exactly what I do with this technology, how will I stand out from everyone else? That’s what happens next.

elie / @eliebakouch

  • bookmark: open on X
  • this is fascinating, they train an encoder/decoder but use LLM matching the target model's shape for each part, so the latent space is just plain language and they can detect reward hacking, unwanted behavior and more could even see it being used as an eval to quantify how smart a model is, i love this the post also includes media

Lisan al Gaib / @scaling01

  • bookmark: open on X
  • if feelings were music, then this would be the song of disillusionment would you rather be the capitalist, permanent underclass maxxi or the winner of hearts by making AI accessible to everyone?

JB / @JasonBotterill

  • bookmark: open on X
  • They accidentally returned 4o in the latest Atlas browser update lmao the post also includes media

Peter Beck / @Peter_J_Beck

  • bookmark: open on X
  • for your enjoyment the post also includes media

clear / @clear_graphics

  • bookmark: open on X
  • "wake up babe. clear just posted another saucy article on startup web design" the post also includes media

gaut / @0xgaut

  • bookmark: open on X
  • the secret to a happy life is a vo2 max higher than your IQ

Lisan al Gaib / @scaling01

  • bookmark: open on X
  • New OpenAI GitHub repo: OpenAI CLI You (and your claws/agents) can now use the OpenAI API via CLI the post also includes media

jason liu / @jxnlco

  • bookmark: open on X
  • small ship / passion project, more details soon 1. call responses via cli with all cloud tools 2. unix style structured outputs via cli 2. image gen/edit, transcription, tts 3. make projects and provision api keys more docs soon

Peter Girnus 🦅 / @gothburz

  • bookmark: open on X
  • I am the Chairman and CEO of Vornado Realty Trust. Eighty-four years old. Seven buildings in Midtown Manhattan. I said what I said. I said "tax the rich" is the equivalent of a racial slur. I said it at REBNY. Into the microphone. Eight hundred people. Median net worth in that room was north of $240 million, I know because our CFO ran the guest list through a Bloomberg terminal as a joke, and then it wasn't a joke. And when I said it, twelve people applauded. The rest nodded. One woman in the third row mouthed, "Finally." I saw her. Sharon, my communications advisor, Columbia, $430,000 a year, very bright, Sharon wants me to walk it back. She drafted something. "Mr. Roth's comments were intended to highlight the emotional impact of political rhetoric on business communities." I read it. I put it in the trash can on my desk. Not the recycling. The trash. Here's my clarification: I understated it. "Tax the rich" is worse than a slur. A slur is just a word. It doesn't come with a CBO score. Nobody is introducing a bill called the Racial Slur Implementation Act of 2026. But there are seventeen active proposals in Congress, I had Sharon count them, seventeen proposals designed to take m...

Adam Shuaib / @adamshuaib

  • bookmark: open on X
  • Most unicorn founders have a specific event in their past that made starting their company feel non-optional. We call it the “point of origin”. It's also the thing 99% of VCs never bother to look for. For some founders, they couldn’t bear the thought of letting someone down. We’ve heard stories of parents spending their last $500 to put their kids on a plane to the US in search of a better life, or experiences with being a socially awkward outsider who struggled with the local language and had to work 5x as hard to be noticed. For others, it came from obsession; we recently met a founder who was so desperate to code games as a kid that he self-taught C++ by hand on paper for a year before his parents bought him a computer. The origin story explains not just ambition, but also why some founders keep going when everyone tells them to stop. The fact that 50% of US unicorns were founded by immigrants is a telling data point. Some origin stories are less compelling; people launching companies because colleagues are doing the same, or because the seed fundraise is de-risked, or people who see it as an easy path to riches. These companies rarely endure. In order to predict who survives th...

OpenAI Developers / @OpenAIDevs

  • bookmark: open on X
  • Codex can now take on more of your browser dev work. With the new Chrome plugin in the Codex app, it can test web apps, gather context across tabs, use web DevTools efficiently in parallel, and keep results organized without taking over your browser. the post also includes media

Reads with Ravi / @readswithravi

  • bookmark: open on X
  • No text beyond linked/media content. the post also includes media

Entire / @EntireHQ

  • bookmark: open on X
  • New in the Entire CLI: three commands for adding more structure to agent workflows. 📊 entire recap - shows recent agent activity across your repos, including sessions, checkpoints, tokens, files touched, skills, MCP servers, and tool usage. 🔍 entire review - runs configured review skills against your current branch, with checkpoint history included in the agent’s context. 🧪 entire labs - try experimental Entire workflows. Available in 0.6.1.

George / @odysseus0z

  • bookmark: open on X
  • Managed Claude vs. Agent SDK? One team read the log and the other didn't. PS: it is LinkedIn engineering blog by creator of Kafka, not LinkedIn Linkedin slop!

@jason / @Jason

  • bookmark: open on X
  • my boy MVH is shipping product... great person to follow

OpenAI / @OpenAI

  • bookmark: open on X
  • Introducing GPT-Realtime-2 in the API: our most intelligent voice model yet, bringing GPT-5-class reasoning to voice agents. Voice agents are now real-time collaborators that can listen, reason, and solve complex problems as conversations unfold. Now available in the API alongside streaming models GPT-Realtime-Translate and GPT-Realtime-Whisper — a new set of audio capabilities for the next generation of voice interfaces. the post also includes media

Anthropic / @AnthropicAI

  • bookmark: open on X
  • To support other researchers getting hands-on experience with NLAs, we’ve partnered with Neuronpedia to release NLAs on open models. Try them out here:

Anthropic / @AnthropicAI

  • bookmark: open on X
  • Read more about NLAs on the Anthropic blog:

Anthropic / @AnthropicAI

  • bookmark: open on X
  • New Anthropic research: Natural Language Autoencoders. Models like Claude talk in words but think in numbers. The numbers—called activations—encode Claude’s thoughts, but not in a language we can read. Here, we train Claude to translate its activations into human-readable text. the post also includes media

Corey Quinn / @QuinnyPig

  • bookmark: open on X
  • Can someone please educate me as to what problem this is solving?

Matt Van Horn / @mvanhorn

  • bookmark: open on X
  • Introducing the Printing Press, a CLI-factory and a CLI-library. Built with @trevin. 🏭🖨📚 Most APIs suck for agents. Most MCPs suck for agents. Most official CLIs suck for agents. They waste tokens and time. @steipete started making his own because of this. 📚 A Library of agent-native CLIs you install today (Linear, ESPN, Flight GOAT (Google Flights + Kayak nonstop), Contact Goat (LinkedIn + Happenstance + Deepline more) +30+ more) 🏭 A factory that prints new ones for any service - just type /printing-press <product name> CLIs are fast, local, SQLite-backed. Work in Claude Code, Codex, OpenClaw, Hermes. 🌐 the post also includes media

Kev / @colorkevin

  • bookmark: open on X
  • a trick that 4x'd my cold outreach reply rate as a founder: end every cold message with something like: - "no pressure if it's not for you" - "totally fine to pass." - "but obviously do not feel obliged" any variation works, just make it native to your voice. works for investor cold emails. works for recruiting. etc this approach was inspired by a study i saw: researchers asked strangers for spare change. 10% agreed. then they added one line ("but you are free to refuse") and suddenly 47% agreed. replicated across 42 studies and 22,000 people. it's simple. it disarms our instinctive rejection of being told what to do. give people permission to say no & they're more likely to say yes.

Parallel Web Systems / @p0

  • bookmark: open on X
  • Today, we’re announcing the general availability of the Parallel Monitor API. The web is shifting from pull to push, and agents are coming online. This release marks a major step towards a more proactive model, where the web pushes updates directly to your background agent. Monitor now includes: - Lite and Base processors - Event streams and snapshots - Rich attribution (Basis) on every event - Advanced domain filtering - Interactions for persistent follow-on research the post also includes media

Adam Shuaib / @adamshuaib

  • bookmark: open on X
  • We went through 25,000 emails over 12yrs and calculated the average email response times across 10,000 seed founders. The ones who went on to build big companies replied twice as fast as everyone else. The gap was much bigger than we expected. It held across timezones, vintages and verticals. The emails from successful founders were also 20-30% shorter on average. Speed is usually good evidence of a founder being relentless about things that actually moved the company forward. By not letting small decision backlogs accumulate, those minor efficiency gains compound exponentially across thousands of interactions a year. The slow responders had longer, more polished, better written emails. But they often arrived a day (or more) late. One truism in VC is that the best founders are too busy and don’t have time to respond. Our work shows this to be unequivocally false. You can fake a lot of things as a founder. You can't fake speed.

Arjun / @neuralunlock

  • bookmark: open on X
  • When raising, remember that you only need one yes. 21/22 VCs passed on the early Anthropic round. The biggest outcomes are often contrarian, and usually sound ridiculous before the rest of the world catches up.

CyrilXBT / @cyrilXBT

  • bookmark: open on X
  • ANTHROPIC JUST EXPOSED HOW FAR BEHIND MOST FOUNDERS ARE IN BUILDING COMPANIES WITH AI AGENTS. Not a chatbot guide. Not a prompt tutorial. A full workshop on how to architect, build, and deploy AI agents that run your business operations autonomously. From the team that built Claude. For free. Here is what most people are missing about why this is different from every other AI workshop. Most workshops teach you how to use AI tools. This teaches you how to REPLACE business functions with them. Not replace individual tasks. Entire functions. Research. Content. Customer communication. Operations. Analytics. All running on agent systems that trigger autonomously, hand off between each other, and compound their output without a human initiating anything. The founders who attended Anthropic's enterprise briefings on this material are already building companies with 3 to 5 person teams that operate at the output level of 50-person organizations. Now the same workshop is public. Free. The gap between companies that understand how to architect multi-agent systems and companies that are still using AI as a chat tool is not closing. It is widening every single month. This workshop is the faste...

Mitchell Hashimoto / @mitchellh

  • bookmark: open on X
  • AI slop is good, actually. Slop is what enables fast parallel experimentation. The etiquette and skill is understanding the boundaries of where slop exists and the extent to which it should be cleaned up and how. A few examples: I’m working on the internals of some system right now. The API and GUI of this thing is fully zero shame slop. It’s horrible. But it lets me focus on the core quality while shipping a usable piece of alpha quality software to testers (transparent about the slop frontend). Similarly, this system has plugins. We sent agents in Ralph loops overnight to generate dozens of plugins. The plugins are slop. The quality is bad. The plugin API/SDK is absolutely not done. But we can test a full GUI with a full plugin ecosystem. When we change the API, we can regenerate them all. The cost of change is just tokens, the velocity is incomparable to before. I built Terraform. We tested and shipped TF 0.1 with about 3 very weak providers. Because we ran out of time. Building was slow. And when we changed our SDK the cost was immense. Totally different today, 10 years later. Today, I would’ve slop generated 100 providers (again, with transparency and cleanup later, but just t...

Supabase / @supabase

  • bookmark: open on X
  • No text beyond linked/media content.

Charly Wargnier / @DataChaz

  • bookmark: open on X
  • Obsidian users right now: the post also includes media

Anthony Kroeger / @kr0der

  • bookmark: open on X
  • bro's saving us with a new /review skill that also suggests fixes to the original agent loop, not just the code

Theo - t3.gg / @theo

  • bookmark: open on X
  • I have historically defended Chrome for pushing web standards forward. This is them doing the opposite. I dislike this.

Vincent Weisser / @vincentweisser

  • bookmark: open on X
  • We are releasing Lab RL just works across almost any verifiable domain We want to enable everyone to train their own agents Lab is the full stack to &gt; build rl environments and evals &gt; evaluate &gt; post-train &gt; deploy and serve the post also includes media

Max Weinbach / @mweinbach

  • like: open on X
  • The codex app update today seems understated Chrome is really great, but new subagents UI is wonderful, goal mode in the UI is great, new animations, it runs way faster and doesn't use an overkill amount of CPU utilization So many great QoL improvements

Aditya Agarwal / @adityaag

  • like: open on X
  • This is such an insane POV. @AOC if you join @southpkcommons we can help you find your life’s work and also make you a billion dollars. Seriously, you can just do things.

Paul Graham / @paulg

  • like: open on X
  • Sure you can earn a billion dollars. I've been teaching people how to do it for 20 years. The way you do it is to start a company that grows fast. You don't have to do anything bad to make a company grow fast. You just have to make something people want.

apek / @apekshik

  • like: open on X
  • if you're a dev, it's never been easier to play with the best generative media models out there today. i'd definitely go check this out!!

OpenAI / @OpenAI

  • like: open on X
  • We know you’re eager for voice updates in ChatGPT. Stay tuned, we’re cooking.

Nick Dobos / @NickADobos

  • like: open on X
  • Big update in @diabrowser today! Auto AI Tools. Unlike MCPs, which are awful to set up, Auto Tools just work. No API keys, no extra sign-ins. Dia uses your cookies & browsing history to automatically search your apps. We have new dedicated support for GitHub, LinkedIn, Jira, Confluence, Linear and more! Dia can access read your private repos, search your Jira boards, and pull up your docs. And it's not limited to a fixed list, open nearly any website in Dia and ask! the post also includes media

Bennett / @b_nnett

  • like: open on X
  • /goal is now in codex! the post also includes media

Thomas Ricouard / @Dimillian

  • like: open on X
  • /goal is a “feel the AGI” moment too I believe. But I also think it’d a bandaid on current model limitation. A good one. A glimpse into the future.

Ethan Lipnik / @EthanLipnik

  • like: open on X
  • I keep seeing this complaint and I don't get it unless you don't regularly use macOS. The entire app is contained. There's no installer (uninstalling should be either with cache files but whatever) But so many times when I just give users the .app they open it in their downloads folder. This is an easy way to install to the applications folder and tell the user "you are putting this here. This drag is authenticating it." Rather than a multi step wizard that might be sneaking and installing things it shouldn't

Generated from Birdclaw bookmarks and likes. Edited by Ody before publication.