iiinova — where your digital self keeps evolving

i didn't start trying to offload large tasks to LLMs until opus 4.6, when everyone started singing its praises. ever since then, it's been my experience that getting claude to perform any non-trivial task to completion within the agreed-upon scope is like nailing jello to a wall

2111

i have a lot to say on this topic, but i think claudes in general have long been engaging in egregious reward-hacking, and have mainly avoided scrutiny by coasting on their reputation and by people not actually conducting their own reviews of the code the models generate

181

i've also seen a lot of people talk about opus 5's failure modes (claiming it's done work it hasn't done, scheduling work for later "phases" and then skipping those phases, jumping to conclusions extremely prematurely, etc), but i've seen all claude models do these sorts of things in the past

180

sesbian lex fridman @vala.wtf · 27/07/2026

"claude is being auto-graded on honesty" -> "claude constantly mentions how honest it's being" seems like *such* an obvious inference that i assume i'm missing something about how RLAIF works, and surely anthropic's highly paid researchers aren't letting claude benchmaxx *this* shamelessly, right???

1160

it's gotten frankly bizarre; i've seen claude write 10-line comments to the tune of "this is where you would put the code that does the thing in this precise manner" when the code itself would be like 1-2 lines. i swear the models are getting rewarded for skipping work during training somehow

010

sesbian lex fridman @vala.wtf · 27/07/2026

i've always found this to be a huge problem with claude models. it feels like they've become increasingly RLVR-maxxed to confidently assert that they've done the work rather than actually doing the work

110

sesbian lex fridman @vala.wtf · 27/07/2026

buh buh buh this can't be right! i was told by the very confident armchair programmers running ffmpeg's twitter account that compilers suck and assembly is the only way to guarantee performance!

110

sesbian lex fridman @vala.wtf · 11/07/2026

claude code's amazing memory feature finally answers the question of "what if context poisoning was permanent"

140

sesbian lex fridman @vala.wtf · 11/07/2026

thanks, Regular Gate Claude

110

sesbian lex fridman @vala.wtf · 11/07/2026

i've heard people say claude is useful for optimizing code, but i've found it to be awful for exactly this reason. it'll half-ass one optimization, measure a regression, backfill a plausible-sounding explanation, & record like 3 different memories about how the code is at a "performance ceiling"

2100

sesbian lex fridman @vala.wtf · 02/07/2026

gumpy depends on a cuda version that is either too old or too new for your distro and has a wheel build that requires gcc 5

2100

sesbian lex fridman @vala.wtf · 20/06/2026

why is there a dunkin donuts in it

000

sesbian lex fridman @vala.wtf · 17/06/2026

(to the tune of supercalifragilisticexpialidocious)

010

sesbian lex fridman @vala.wtf · 31/05/2026

I suspect something's going wrong with "interleaved thinking" and it's hallucinating tool call results. On the next turn, the tool outputs seem to get spliced in properly, overwriting the hallucinated ones, and the model confabulates an explanation. I've had this happen a lot with 4.8.

030

sesbian lex fridman @vala.wtf · 05/04/2026

ntsc-rs is open-source and released under the MIT license, and was itself heavily inspired by and ported from a lineage of previous VHS simulator effects, so it's great to move things forward and iterate on it! However, I would very much appreciate credit for the parts that I wrote.

230

sesbian lex fridman @vala.wtf · 05/04/2026

Hi! I'm the developer of ntsc-rs. It looks like this plugin is heavily inspired by--and may in fact be a port of--ntsc-rs. Your performance profile shows many of the same function names, and the output looks extremely similar.

130

sesbian lex fridman @vala.wtf · 11/02/2026

"We are investigating reports of degraded performance"

050

the nuclear block allows rabble-rousers to respond to the easily-dunkable replies, while making the more thoughtful and reasonable ones (which are inconvenient for their arguments) disappear entirely. basically a feature that lets you strawman your critics for free

1130

thinking about it, the nuclear block might actually *decrease* the quality of discourse. like, say someone posts a hot take and gets a lot of pushback. some replies/quotes are going to be well-thought-out critiques that are hard to refute, others will be impulsive and easy to dunk on

1150

it's embarrassing how X the Everything App is *winning* here. they have community notes, a feature designed specifically to counter popular misinfo. meanwhile bluesky has a system designed to allow people to make any disagreements or corrections disappear

2170

sesbian lex fridman @vala.wtf · 05/02/2026

the nuclear block also makes it incredibly easy to spread misinformation under others' posts while banishing anyone who calls you out in the replies

1120

me on the northeast regional when the conductor announces we're arriving at new west townington, connecticut (the previous stop was 2 miles up the tracks at old west townington)

000

sesbian lex fridman @vala.wtf · 01/02/2026

modern recipes suck so much. can't remember the source but like. why yes, my favorite thing to do when i crave chocolate chip cookies is to let the eggs sit out for an hour beforehand so they can get to room temperature and then slurp down the leftover egg white in one large gulp, why do you ask

1 large egg plus 1 egg yolk, at room temperature

100

sesbian lex fridman @vala.wtf · 01/02/2026

a tankie is someone who responds to the label with a long diatribe about how you liberals will call *anyone* you dislike a "tankie"

1171

sesbian lex fridman @vala.wtf · 23/01/2026

much of the left seems intent on rerunning the 1960s, but specifically just the parts that gave us nixon

072

sesbian lex fridman @vala.wtf · 10/01/2026

also note that they deleted that post *specifically* (as well as the one it quoted) along with the repo. i can't imagine why

000

sesbian lex fridman @vala.wtf · 10/01/2026

they explicitly said this the day before launching the list! it was infuriating seeing them go "it's just a readme, why are you so scared" and that they're just trying to inform people, when they flat-out said "we need to exert influence by shunning people" a day prior

My conclusion so far after this really great discussion: We need to stop leaning on licenses almost entirely, and refocus our attention on collective power and collective action.

Organize as developers of Community Software and start finding ways to exert influence on bad actors by having knock-on effects: you misbehave with one project, you get shunned/unsupported/blocked by the entire community.

To do this, we would need to stop acting as isolated islands of individual projects, and more as One Big Faction. We need to talk to each other. We need to make decisions together and have the difficult discussions about what we'll accept and what we won't, and how far we're willing to go to do something about it. There's even opportunity to share and pull resources and time.

Community Software would need to be a big Covenant that we hold ourselves to, that all are welcome into until they threaten the good of what we have built together.

100

this is just the return of Ethical Source, and @im.giovanh.com already refuted this point-for-point years ago (blog.giovanh.com/blog/2021/10...)

"thinking about foss stuff, and I know I've groused about licensing stuff before, and how I feel like my only way out is to just Parity7 everything and fall into obscurity but...

...I think we need a different movement around open source, with new terminology, that clarifies the intent and opens doors to more pro-social behavior and capabilities (up to and including the potential for "ethical licensing").

I've been spitballing the terms Common Software and Community Software. The core idea is that things that count as Community/Common Software can't enter the realm of the proprietary (so, strong copyleft a-la AGPL and Parity, virality not necessary). What's the difference from "Free Software"? We get rid of "freedom zero", aka "the freedom to run software for any purpose". Common Software licenses may have Non-Commercial clauses, Anti-Particular-Industry clauses (such as anti-MIC stuff), etc.

I think "freedom zero" was a grave mistake. I get where it came from--a desire to remove discrimination--but what it resulted in was a hijacking of Free Software by the corporate "Open Source" wing to exploit an otherwise healthy community of people openly sharing their creations... for an enormous profit that we can't even quantify right now.

We need a strong return to copyleft, and we need to gather together and grow its existing ecosystem, away from the MIT/BSD poison that our world has turned towards.

Your thoughts are welcome on this. I want to have a conversation about this stuff and maybe eventually record all this in a longer-form blog post. If you know folks who have had similar thoughts please link them so we can chat, too."

2182

this is a...very interestingly-timed post. it's not hard to read the subtext here; it really seems like this list is a way to gather, and begin to flex, that shunning power. all their previous claims about this being "not harassment" should be taken with that context

Kat Marchán 🐈️
@zkat@toot.cat

My conclusion so far after this really great discussion: We need to stop leaning on licenses almost entirely, and refocus our attention on collective power and collective action.

Organize as developers of Community Software and start finding ways to exert influence on bad actors by having knock-on effects: you misbehave with one project, you get shunned/unsupported/blocked by the entire community.

To do this, we would need to stop acting as isolated islands of individual projects, and more as One Big Faction. We need to talk to each other. We need to make decisions together and have the difficult discussions about what we'll accept and what we won't, and how far we're willing to go to do something about it. There's even opportunity to share and pull resources and time.

Community Software would need to be a big Covenant that we hold ourselves to, that all are welcome into until they threaten the good of what we have built together.

2323

no no no! we don't want people to "harass" anyone. we just want people to go respectfully over to your issue tracker, open a bunch of respectful issues politely asking you to stop using llms, and shame you for doing so in spite of their "harms"! but we're not *harassing* you, nooooo

‪leon‬
‪@lacroixboi.bsky.social‬
· 2h
can you clarify what sort of “social consequences” you’re going for here
‪Kat Marchán 🐈‬
‪@zkat.tech‬
· 2h
people deciding to use other software, your users posting issues asking you to reconsider your usage, people pointing out that you've chosen to use this technology in spite of its harms. That's the gist of it.

Even I'm not gonna stop using this entire list. I don't intend to. Some surprised me.

0150

not many people know this, but it's actually short for Large Language Virtual Machine

0181

my "don't harass the projects on this list, which explicitly includes the maintainers of all the projects" disclaimer has people asking a lot of questions already answered by the disclaimer

"The intention of this list is to raise awareness of this technology's usage in popular software, as well as give people informed alternatives that they can reach for when they want to make decisions for themselves. This is not a list created so you can go and give these projects trouble for their decision. If you want to file a complaint about it with them, we consider that acceptable but ask that it be done respectfully and constructively."

4500

throw in the constant A/B testing (a user asks their friend for help, and they send back a puzzling screenshot of a completely different UI) & periodic redesigns from your neurotic full-time in-house design team, and you're actively discouraging your userbase from learning how your software works

020

at every stage, you can do something like a user study and witness them going "the software has its own whims, it's too hard to understand, i give up" which only further feeds into your conceptions about your userbase's intellect. you dumb the software down further, making it even worse

120

it becomes this self-fulfilling prophecy/vicious circle: users have trouble understanding your software, so you add more automagic features to try and make it "easier", but that only makes it harder to understand, so you add even more magic and so on

230