weekly recap #1: apr 20 - apr 26, 2026

one thing i am noticing with working with agents is that there is much more mental fatigue after a day of work compared to before. i feel like i am spending less time within a single context. prior, i had 1-2 main context windows per day, but now i am juggling 5-10+. it puts a cognitive strain that is hard to shake, this constant feeling of being stretched too thin. if i could just spawn more of myself like i spawn more agents, everything would be fine.

but it raises a question: because we can be more productive, should we? should the tools be used so we can produce more, or should they be used so we can work less and spend our lives on things other than work? because we have washing machines, should we buy more clothes so we can wash them more often? are we being more productive out of necessity or are we inventing things to do?

i dont know the answer. but i feel like something is wrong.

• written after a long day of working with agents. tiring, but nonetheless productive

what i read this week

• openai's bet on compute
  • openai gigantic bet that compute is how you scale the models (scaling referring to model intelligence, not handling "load"). interesting to think about as anthropic is struggling from success, not being able to provide enough compute to keep up with demand. tweet to the point where their new opus 4.7 model uses 33% more usage than opus 4.6. at relay we hit our pre-opus 4.7 claude limits within 1-2 hours of it being released.
  • right after writing this, i opened up twitter to see that amazon and anthropic are working together to provide more compute to anthropic, reportedly 5GW of datacenter compute capacity. (for reference, that is more than all of microsofts datacenters compute capacity in 2024).
• how i use amp
  • a bit old, but nice points of how an amp team member uses amp
• typing can be the bottleneck
  • i think out of all of the new tools the AI advancements have enabled for us is tools that help with dictation. for example, i have been using and loving wispr recently. now pretty much whenever i am writing a prompt, a project update, a long slack message, i use it. it boosts my WPM and lets me spend more time thinking instead of typing. also, i find that speaking an idea out loud helps me find holes in it :p
• kelpdao suffered an exploit for $290m rsETH, allegedly due to a forged cross-chain message. initially seems to be a sophisticated attack originating from lazarus, but seems to be a consequence of poor security practices around DVNs (1-of-1 setup). it then lead to a cascade of bad-debt rumors for aave as rsETH was being used for collateral
  • kelpdao response
  • layerzero response
  • aave response
  • momentous moment as arbitrum decided to freeze $76m worth of eth related to the attack
  • bad month for defi, as earlier this month drift protocol also got hacked by north korean hackers for $285m
• a third party vercel tool/oauth app also got hacked, with hackers allegedly getting access to vercel users' env variables. not super clear what happened here, as vercel says they were able to get "unauthorized access to certain internal Vercel systems" vercel security incident
• lovable also had a bug that caused a data breach where anyone can look up other people's projects. including prompts, source code, etc. apparently it was reported in november of 2025 originally, but lovable decided to only fix it for new projects, with old ones still impacted. user here was able to demo how they extracted info from another user's project.
  • lovable responded by claiming that they did not suffer a data breach, and instead, it was made unclear to users the public nature of their projects: "Our documentation of what 'public' implies was unclear, and that's a failure on us."
• someone offering a 14 acres estate in Mill Valley california for Anthropic stock, but the seller also retaining 20% of the stocks upside at the time of stock sale. didn't know people did things like this
• John Ternus is a safe pick.
• USVC, regulated fund that lets retail invest in private companies
• a mev guy accidentally exploited Kipseli for 80k cbbtc, then later returns 80% of the funds in an onchain series of messages between kipseli and the exploiter. must read
• providers overquoting
  • enso shield
• this described burnout perfectly. whenever i feel burnt out, a break does nothing except worsen it. it feels like every time to get past it i need to rethink what i want and what my goals are, both short and long term. great post
• proof you can do hard things

  • Our self-image is composed of historical evidence of our abilities. The more hard things you push yourself to do, the more competent you will see yourself to be. If you can run marathons or throw double your body weight over your head, the sleep deprivation from a newborn is only a mild irritant. If you can excel at organic chemistry or econometrics, onboarding for a new finance job will be a breeze.

  • reminded me of something i wrote before, on getting better > you don't just learn new things, you also learn how to learn new things. and every time you pick up a hard project and complete it, you raise the bar on what you are capable of.
• i always look for curated lists of essays, writing, etc. especially by those that write things i like to read. one of my go-tos is masonjwang's bookshelf (an online bookshelf is really cool concept too i might copy :) )
• bankrate checking accounts survey

  • Account holders have had the same checking account for an average of 19 years, 17 years for savings

• github had a really bad incident where previously merged commits were getting reverted by proceeding merges
• twindex, an interesting experiment to have two agents (in this case, codex) pair program together. one agent acts as the engineer that implements code, the other as the navigator that approves/instructs it to do things. the idea is that the navigator replaces the human.
• patrick collison's advice
  • my favorite quote from this: "To the extent that you enjoy working hard, do."
• i learned about writing "hooks" in essays over and over again during school and never gave it much thought. i love how gwern explains it here
• opus 4.7, the new smart mode in amp
• how handoff was built in amp
  • really cool insight into how handoff is built (something i use really frequently with amp), and insight into how building amp works. it also brings to light how important the opinions of the people who build amp are on how good it is. there are times i used tools that use the same models amp uses, but it just felt shallow, dumb, and not as sharp as amp. this post helped show how much thought goes into a feature like handoff.
• kirby is eating context engineering talks about how context engineering is becoming swallowed by newer more advanced models, that know to do things to preserve context without being instructed to do so, like using subagents, efficient tool use, even researching and planning.