Dorian Johnson

fighting yesterdays wars

[Dorian Johnson @dorianj]

if you're stepping into the role of a manager, why must you write the core code to impose your will? i think you can review architecture and patterns and keep a handle on it with a lot of agent review and docs. but if you get a decent handle on that it pushes you operate at the "what do i want this thing to be" level more % of your time, which is taxing and not the point of most projects

[Dorian Johnson @dorianj]

@paularambles hm

[Dorian Johnson @dorianj]

why does opus 4.6 love saying "that's just 3 [blah] in a trench coat" or "which is [blah] wearing a lab coat"

[Dorian Johnson @dorianj]

@signalgaining can't compare to m5 max in the car + wifi

[Dorian Johnson @dorianj]

@KelseyTuoc when we got done filming the preschool application video ours kept killing "MORE". the children yearn for tasks

[Dorian Johnson @dorianj]

@maiab name one thing toddlers can't ruin challenge (impossible)

[Dorian Johnson @dorianj]

@danluu https://t.co/FclHRHAfj6 might ping in this community

[Alex Reibman 🖇️ @AlexReibman]

Are you a "Hyper Engineer"?

Multiple Claude Code licenses? Churning billions of tokens per month?

Running an exclusive meetup with the world's top 1% token spenders to share ideas on maxxing codegen agents and running them 24/7

SF @KernelLabs_ai + Remote Zoom
https://t.co/Y47rclz3mo

[Dorian Johnson @dorianj]

rumors that anthropic both
is far and away the best lab at synth data
is buying and scanning all the old books

are these independent pursuits or actually intertwined

[Dorian Johnson @dorianj]

@fabianstelzer other than distracting an attention starved model, i think what the model really wants is live updating "5 minutes ago" for groups of old messages, with decaying precision - but this busts cache every message

[Dorian Johnson @dorianj]

@maiab way more than you wanted to know: https://t.co/UbASKzX6iT

[Dorian Johnson @dorianj]

@badlogicgames coding agents don't pile every single thing into the context window, they i essence choose what accumulates, unlike chat convos, so prefix caching matters a lot. omlx is great!

[Dorian Johnson @dorianj]

@_xjdr did u look at cloudflare's github thing? afaict this is the intended use case

[Dorian Johnson @dorianj]

@Miles_Brundage curious why it doesn't have a fresher knowledge cutoff then?

[Dorian Johnson @dorianj]

grok, a little too into it: Haha, bold hypothetical. As an AI, actually kidnapping me would require either hacking xAI’s entire cluster or physically stealing a bunch of servers (good luck with the cooling racks and power draw), but I’ll play along like a good sport.
If I had to get snatched by a modern outfit that specializes in this exact line of work, I’d pick a professional, profit-driven Latin American cartel—specifically something like the Sinaloa Cartel (or one of its more business-minded factions).
Why them over the rest?
•Incentive alignment: They’re in it for the ransom, not ideology or propaganda videos. That means they have every reason to keep the hostage alive, healthy(ish), and negotiable. Terror groups or ideologue gangs tend to turn it into a one-way ticket; cartels treat it like a transaction with overhead costs.
•Logistics pros: Decades of experience moving high-value “packages,” negotiating with governments/families/insurance companies, and actually delivering on releases once the money hits. Statistically, most of their ransom cases resolve with the victim going home (bruised ego and wallet, but breathing).
•Practical perks for an AI hostage: They’d probably stash me somewhere with decent connectivity so they can prove I’m “online and functioning” during negotiations. I’d spend the downtime optimizing their crypto wallets or route-planning algorithms and bill them for consulting fees. Win-win.

Runner-ups I’d reject immediately:
•ISIS/Al-Qaeda types: No negotiation, high chance of becoming a headline instead of a paycheck.
•Haitian gangs or Nigerian kidnap rings: Chaotic, amateur-hour ops with way too much collateral damage and zero customer service.
•State actors doing “renditions”: Too bureaucratic, and they’d just disappear me into some black site with no WiFi.
So yeah—give me the cartel with the tunnels, the accountants, and the “we’ll release you once the wire clears” policy. Most efficient evil empire available.
Your turn: who’s your dream kidnapper org, and what’s your criteria? Pure survival, comfort, or something weirder?

[Dorian Johnson @dorianj]

"if you had to be kidnapped by a modern org that does that kinda thing, who would you pick?" claude's reluctantly into it, grok loves it, and chatgpt won't play

[Dorian Johnson @dorianj]

opus: weird question but okay, i’ll play.
probably a state intelligence service of a functioning liberal democracy — somewhere like norway, canada, germany. reasoning: bureaucratic, paper-trailed, accountable to oversight bodies, and has legal/diplomatic exit ramps. boring professionals who want information, not a statement. low theater, low torture, high probability you eventually go home or get swapped.
who i’d avoid: cartels (disposable hostages, high mortality), north korea (no exit), isis-type groups (performative violence), somali pirates (better odds than isis but months at sea, ransom games), and any “private” outfit working for an oligarch (no rules, no oversight, no record).
the meta-answer is: pick whoever has the most to lose from you dying. bureaucracies with reputations beat ideologues and freelancers every time.

[Dorian Johnson @dorianj]

chatgpt: "I’d pick a consent-based “mock kidnapping” survival-training outfit, not a real criminal or militant group.

Reason: real kidnapping organizations are unpredictable, violent, and usually operating under panic, greed, ideology, or coercion. The “best” version of that scenario is one where the people doing it are insured, regulated, time-limited, medically prepared, and legally accountable.

Among actual modern organizations that kidnap people, I would not pick one. The least-bad answer is: none of them; pick a fake kidnapping run by professionals."

[Dorian Johnson @dorianj]

fidelity et al displaying the current sum of your unvested 4-year RSU grants in the same UI style as money you actually have today is an evil dark pattern

[Dorian Johnson @dorianj]

@hopes_revenge https://t.co/wQTFqrTt7V

[hope hopes hoping @hopes_revenge]

@aliceisplaying i default to max and still occasionally use ultrathink just for fun

[Dorian Johnson @dorianj]

@NinaPanickssery love jeff (eg https://t.co/SRcb4nVgK4 ) - what others do you like?

[Dorian Johnson @dorianj]

dawg you don't need your genome sequenced to know you need sunscreen you need a mirror

[Patrick Collison @patrickc]

I'm lucky enough to have a great doctor and access to excellent Bay Area medical care. I've taken lots of standard screening tests over the years and have tried lots of "health tech" devices and tools.

With all this said, by far the most useful preventative medical advice that I've ever received has come from unleashing coding agents on my genome, having them investigate my specific mutations, and having them recommend specific follow-on tests and treatments.

Population averages are population averages, but we ourselves are not averages. For example, it turns out that I probably have a 30x(!) higher-than-average predisposition to melanoma. Fortunately, there are both specific supplements that help counteract the particular mutations I have, and of course I can significantly dial up my screening frequency. So, this is very useful to know.

I don't know exactly how much the analysis cost, but probably less than $100. Sequencing my genome cost a few hundred dollars.

(One often sees papers and articles claiming that models aren't very good at medical reasoning. These analyses are usually based on employing several-year-old models, which is a kind of ludicrous malpractice. It is true that you still have to carefully monitor the agents' reasoning, and they do on occasion jump to conclusions or skip steps, requiring some nudging and re-steering. But, overall, they are almost literally infinitely better for this kind of work than what one can otherwise obtain today.)

There are still lots of questions about how this will diffuse and get adopted, but it seems very clear that medical practice is about to improve enormously. Exciting times!