chip 🔒 @wd.teleri.net · 20/09/2026the people who get nerfed are obviously pissed when you take their toys away, but they’re both a minority of both the current active player base , and even more so of potential players. you want to work harder to engage future participants! you already hooked the top players, they’re not quitting 020
chip 🔒 @wd.teleri.net · 20/09/2026preventing massive disparity through taxation is just balancing. you want your audience to believe they have a credible path to the top of the leaderboard. in theory capitalism actually should do this through efficient markets etc, but that demands … well, balancing 130
chip 🔒 @wd.teleri.net · 20/09/2026I’m sure you can poke holes in this idea, but if your goal is to maximize productivity / output, then it behooves you to maximize engagement. People won’t engage in clearly rigged games. If the top of the ladder is unreachable, you’re just losing potential participants to apathy / disinterest 120
chip 🔒 @wd.teleri.net · 20/09/2026in gaming terms this is just a nerf. and we should nerf overpowered builds, so to speak. nerfing universally incentivizes people who keep playing to find new ways to succeed and find new ways to run up their score (ie “create economic value”). balance patches are good, taxes are just that 010
chip 🔒 @wd.teleri.net · 20/09/2026not an argument for anthropomorphizing the models in itself, but people will naturally trend towards convenience and efficacy and that is self-reinforcing. I don’t think you can train it out of the models or out of people, basically 010
chip 🔒 @wd.teleri.net · 20/09/2026my experience remains that language models respond more favorably to polite and inclusive language, likely an artifact of training on human discourse. so users are rewarded for engaging with them in that fashion. while efforts have been made to reduce this effect it may never go away fully. 110
chip 🔒 @wd.teleri.net · 20/09/2026this show has been well-dissected by others and I won’t do it justice, but being able to get a 52 episode runway really let them cook in a way you don’t see much of anymore, and that’s a shame. it was a gateway for mythology and literature for kids who got sucked in unlike any of its contemporaries 130
chip 🔒 @wd.teleri.net · 20/09/2026don’t start. it’s a terrible nerd trap. also at some point Disney got involved which didn’t help 020
chip 🔒 @wd.teleri.net · 20/09/2026they’re definitely not universally good. but it does feel like they get extra budget and thought. exception proving the rule though, I’m sure 060
chip 🔒 @wd.teleri.net · 20/09/2026for some reason I find myself watching Gargoyles yet again. this show was too good for its Disney afternoon slot. a ludicrous collection of voice talent, serialized drama in a sea of episodic media. an audacious four episode series open. I don’t know how they got away with it but I’m glad they did 141
chip 🔒 @wd.teleri.net · 17/09/2026Blizzard appears to have COOKED with Diablo IV season 15. I would like to not be on constant escalations/bridges so I can spend more time with this. Claude, rewrite our infrastructure. Make no mistakes. 🫠 030
chip 🔒 @wd.teleri.net · 15/09/2026haha ty, had a lot of stuff happening, miss this place but juggling too many distractors. hopefully I can skeet regularly again … in moderation 010
chip 🔒 @wd.teleri.net · 15/09/2026on the other side I do feel compelled to do more to utilize my 64GB of unified memory on my MBP for certain kinds of work 120
chip 🔒 @wd.teleri.net · 15/09/2026it feels inevitable that more compliance requirements (and expensive auditing) will continue appearing to prove SBOM-but-for-AI for critical sectors. US gov (and plausibly EU tbh) will basically ban Chinese model use for critical work 110
chip 🔒 @wd.teleri.net · 15/09/2026that paid-for lead and cost honestly doesn’t feel unreasonable by default, if you’re doing like line of business apps or you’re an IT shop you can be behind a bit but pay-for-premium is hardly novel and some people likely benefit. cost doesn’t sound outrageous on its face 110
chip 🔒 @wd.teleri.net · 15/09/2026interesting results that are not shocking from this Mozilla study on open weight vs. frontier AI models: arstechnica.com/ai/2026/09/e...arstechnica.comExclusive: Paying for frontier AI models buys 4-month head start at 5x the costArs previewed Mozilla’s report on how cheap open models caught up on capability. 130
chip 🔒 @wd.teleri.net · 06/08/2026this appears to be Windows 95/98/2000-era sensibility. My guess (elsewhere) is that it's a Visual FoxPro app ca. ~1999-2000. Guess mostly on vibes, but that UI screams ~win98 to me if nothing else. 010
chip 🔒 @wd.teleri.net · 14/07/2026while I am at most an armchair physicist, it seems like this kind of capability needs a really absurd amount of power to actually work. the tens of kilometers of atmosphere through which a powerful light must be projected (and through which said light diffuses) seem like a significant hurdle 150
chip 🔒 @wd.teleri.net · 11/06/2026Sonnet 4.6 remains suitable for most one-off inquiries. Which is mostly great from a user perspective, because Sonnet cost keeps dropping proportional to frontier models used for complex specialist work. 010
chip 🔒 @wd.teleri.net · 11/06/2026the beauty of unpacking is there is no schedule. plus if you have a box of stuff you haven’t unpacked for like five years it’s … probably just trash. 040
chip 🔒 @wd.teleri.net · 11/06/2026I think post edits and revision history are probably the right thing for users. There’s a reason I like these things in other platforms. It promotes a looseness that a place like bsky should embrace. But it’s not a no-brainer value-add, even though it feels like it 010
chip 🔒 @wd.teleri.net · 11/06/2026Modern messaging has totally broken my brain, in terms of composition. I’ll forget on IRC that *nobody knows it took me five minutes to compose a reply* because there’s no typing……. indicator. When I do recall this, it’s freeing. 110
chip 🔒 @wd.teleri.net · 11/06/2026Yes, especially when I’m rapid posting and fuck up. However, there’s this part of me that thinks it’s better if we just own our errors and have to correct in a follow up. Using IRC more again has made me cognizant of how much I rely on fast-editing my messages instead of … stopping and thinking. 230
chip 🔒 @wd.teleri.net · 10/06/2026also, oof, this is infact a fork of uutils-coreutils and i simply missed that when i looked at this last week. i remain baffled that someone took the time to bounce `dircolors` from the Microsoft distribution, declaring it "not useful", but hey! 100
chip 🔒 @wd.teleri.net · 10/06/2026it's still great! i'm really happy to see any genuine improvements here. i just wish for the bandaid to be truly ripped off, and to have a Real Shell That Isn't a Miserable UX in Windows instead of a drip-feed of half-measures towards that 100
chip 🔒 @wd.teleri.net · 10/06/2026I am so deeply frustrated that this is a distinct Rust port of coreutils rather than simply improving the (excellent, already extant, and more compatible) uutils-coreutils. it is such a two steps forward/one step back maneuver 100
chip 🔒 @wd.teleri.net · 10/06/2026a take i'm sure i shouldn't post: as part of modern history education, students should be forced to use the internet on a 14400bps link for a couple weeks straight and figure out how to satiate their carnal desires imagine having the bandwidth for *JPEGs*. 000
chip 🔒 @wd.teleri.net · 10/06/2026if you've never seen wdavdaemon your macbook is still living its best life 100
chip 🔒 @wd.teleri.net · 10/06/2026here's the thing ben i'm in the zone, i don't have time to distract myself with micturation, it's going to set me back [some interval of time] that i cannot countenance. 000
chip 🔒 @wd.teleri.net · 10/06/2026i had to give a whole-ass complex template to an older generation of models to explain "here's how you write sphinx w/o types python docstrings the way i want." but newer models can exist with just "sphinx without type decorations" as the directive and perform equivalently. many saved tokens. 010
chip 🔒 @wd.teleri.net · 10/06/2026also, model release cadence continues to be stupid fast. reassess new models with baseline behavior. no accumulated user config, no pile of skills, no honed instructions. try "unburdened by context" tasks with model+1/model+2 revs if you have the time. discard now-superfluous instructions 210
chip 🔒 @wd.teleri.net · 10/06/2026tokens feel analogous for me to low memory in the DOS days. a scarce resource that needs careful consideration and management. prefer to bring tokens in knowingly. if you have a bundle of included instructions on writing great docs, but you're vibing Python code, how are those helping? they're not 120
chip 🔒 @wd.teleri.net · 10/06/2026just from personal experience/trials, i've found that avoiding compaction/compression is broadly positive. the easiest way to do that is to not burden *every new model prompt* with superfluous (and distracting!) instructions, skills, etc that may not be relevant to the task at hand 110
chip 🔒 @wd.teleri.net · 10/06/2026DO: use `<FOO>.md` instructions (claude, agents, whatever) *judiciously*. DO: structure these as discrete skills, or in subfolders, for large projects DO NOT: feel like if you do `/context` and you've blown 25-30% of your token window before you type the first prompt that it's nbd 110
chip 🔒 @wd.teleri.net · 10/06/2026my work continues giving me infinite tokens for, well, work. it "feels good" to use the most powerful available machinery for every task, but as the sudden onset of "tokenomics" is hitting people, for my own personal use on small codebases I genuinely think Opus (or comparable) is mostly overkill 120
chip 🔒 @wd.teleri.net · 10/06/2026your context window is shockingly small, compression remains a pitfall. the compression is lossy, just like most A/V codecs, but the lossiness is much worse for the model than for a human consumer. if you cook 30% of your tokens on ANY NEW PROMPT on a pile of instructions/skills, you're losing out 010
chip 🔒 @wd.teleri.net · 10/06/2026genuinely sonnet 4.6 for internet search, and low-stakes coding in small projects/blitting out toy scripts. opus 4.7/4.8 if you're done caring about tokens for a while or on the turbo $$ subscription. verbose custom prompts and a littany of skills/etc are a terrible trap, be judicious here 210
chip 🔒 @wd.teleri.net · 10/06/2026siri, for me, is not fixed until i can say to my homepod "turn on the dining room lights and set the brightness to 50%" and not have it tell me "bro idk i can't do two things at once." if this is what gets it there, great. i am ... pessimistic that it will happen. probably i'm not the target user 000
chip 🔒 @wd.teleri.net · 10/06/2026i think a lot about a random passage in verner vinge's zones of thought series where code archaeologists find unix timestamps hidden amidst the sublayers of code they normally give a shit about. genuinely prescient writing. 010
chip 🔒 @wd.teleri.net · 10/06/2026I would unironically be into a 45min PFTPM deep dive on this ~next week when the podcast/morning show goes on hiatus and I am otherwise bereft of PFT content. My cynical take is that some amount of "skimming" was an inevitable outcome of the settlement, but I wonder whether this is beyond the pale? 100
chip 🔒 @wd.teleri.net · 10/06/2026if i'm starting fresh i'm using the TUI app in [whatever repo/folder] and sort of describing what i want in a blank canvas, then spawning a boatload of tabs to inspect the output. if i need quick changes to existing code i was already working on, i'll just use the sidebar chat thing 010
chip 🔒 @wd.teleri.net · 10/06/2026work means i use github copilot CLI for most things in lieu of claude code, but i find myself surprisingly bouncing between either of those TUI apps (they're mostly not CLIs, terrible framing imo) and... the VS Code secondary sidebar where the agent stuff lives. and I really viscerally dislike GUIs 110
chip 🔒 @wd.teleri.net · 10/06/2026i'm noodling on a post (or series?) about how software development has been upended for me in the last ~12-18 months, in the interim i spent too much time analyzing a favorite weird al song etc.teleri.net/2026/05/16/b...etc.teleri.netThe Chronology of The Biggest Ball of Twine in MinnesotaThe Chronology of The Biggest Ball of Twine in Minnesota Been a ridiculous amount of time since I posted anything on my blog, so instead of talking about my work, the state of AI coming from deep in t... 120