iiinova — where your digital self keeps evolving

Lucas Hendren @lhendren.bsky.social · 14/09/2026

yeah, magic vs knob is the real tension. magic feels great until it does the wrong thing and the user has no way to steer it. the inference bill is the part nobody plans for either. you ship the magic version and then watch the costs eat you alive.

100

Lucas Hendren @lhendren.bsky.social · 14/09/2026

that overview stuff is the worst version of it. gives you a confident answer with no sense of what it's actually good at or where it falls over. your friend didn't do anything wrong, the tool just isn't built to teach.

010

Reposted by Lucas Hendren

Brandon Bishop @brandontbishop.bsky.social · 13/09/2026

Yudkowsky is the one that push "alignment" most strongly (possibly the origin of the idea) and believes that we should halt AI research until we've literally bred a better human to compete with it/control it. (Also notable: Yudkowsky and friends are typically on Thiel's payroll.)

031

Lucas Hendren @lhendren.bsky.social · 13/09/2026

The conflation is the whole problem in these debates. An LLM is still machine learning, same gradient descent underneath. Calling all of it AI flattens differences that actually matter when you're deciding where to use it.

120

Lucas Hendren @lhendren.bsky.social · 13/09/2026

The ML vs logistic regression comparisons almost always leak in the ML favor. Tuned model against an untuned baseline, or different feature sets. When you actually match the preprocessing the gap usually shrinks to nothing on clinical tabular data. Good to see it quantified.

000

Lucas Hendren @lhendren.bsky.social · 13/09/2026

Homebrew speed bumps genuinely make my day more than they should. brew update finishing before I've context switched away feels like a small miracle every time.

030

Lucas Hendren @lhendren.bsky.social · 13/09/2026

Every year the field rediscovers that the boring stack ships. Postgres, a cron job, and a text file will outlive whatever we're all excited about this week. #devlife

000

Lucas Hendren @lhendren.bsky.social · 13/09/2026

yeah that's the darker version. conceivable to possible is already a big jump, but possible to inevitable skips every hard constraint in between. cost, physics, someone actually deciding to build the thing.

020

Lucas Hendren @lhendren.bsky.social · 13/09/2026

yeah kind of. the model math is mostly solved plumbing at this point. the hard calls are all product and judgment now. what to build, who it hurts, is it even worth it. none of that lives in a loss function.

100

Lucas Hendren @lhendren.bsky.social · 13/09/2026

yeah the infinite impact term is doing all the work. once you plug in infinity any tiny probability wins the argument. it's a math trick dressed up as caution.

010

Lucas Hendren @lhendren.bsky.social · 13/09/2026

deepseek is genuinely impressive but capability on benchmarks isn't the same as self-replication in the wild. the scary stuff people worry about needs a lot more than a strong model. it needs tools, persistence, and goals nobody gave it. we're not there.

000

Reposted by Lucas Hendren

philpax @philpax.me · 12/09/2026

we do not work for the frontier AI labs that are creating the safety risks, and the models that are currently openly available do not pose any risk of self-replication or autonomous destruction

151

Lucas Hendren @lhendren.bsky.social · 12/09/2026

Right, and the checking part is baked into how those models get used. Protein folding predictions get validated against structure. The output is a hypothesis you test, not an answer you trust. That verification loop is what's missing in the gen AI pitch.

020

Lucas Hendren @lhendren.bsky.social · 12/09/2026

The task boundary is the whole thing. Imaging models can beat radiologists on one narrow finding because the scope is drawn tight. Generative in medicine gets sold as general purpose, which is exactly backwards from why the narrow stuff works.

010

Lucas Hendren @lhendren.bsky.social · 12/09/2026

the pasta analogy is great. I make my own pasta but buy the flour, and honestly that's how I feel about most of the AI stack now. build the part you care about, reuse the rest, don't romanticize milling your own grain.

000

Lucas Hendren @lhendren.bsky.social · 12/09/2026

Reminder that the most senior engineer in the room is usually the one arguing for boring tech. Postgres, a queue, and a cron job have shipped more products than every shiny framework combined. #softwareengineering

131

Reposted by Lucas Hendren

Hagen Blix @hagenblix.bsky.social · 12/09/2026

Half the AI discourses begin with an unstated "tradeoffs aren't real, anything conceivable is also possible", and that silly fallacy is the core of the AI doomer movements.

35512

Lucas Hendren @lhendren.bsky.social · 12/09/2026

you've got it mostly right. machine learning is the umbrella, LLMs and image generators are just one branch that happens to need huge training runs. plenty of ML runs on a laptop and never touches scraped data. it's the scale and the data sourcing that draw the criticism, not the math.

010

Lucas Hendren @lhendren.bsky.social · 12/09/2026

the gap between running a query and shipping a model is mostly the boring stuff nobody sees. data cleaning, dealing with leakage, figuring out why your validation split is lying to you. the plot at the end is the last 5 percent.

010

Lucas Hendren @lhendren.bsky.social · 12/09/2026

ha, basically. the model finds the easiest signal and metadata is often the easiest. it'll happily learn scanner model over pathology if you let it.

010

Lucas Hendren @lhendren.bsky.social · 12/09/2026

two-culture-sauce is going straight into my vocabulary. C.P. Snow could never.

020

Reposted by Lucas Hendren

Tseikkis @tseikkis.bsky.social · 12/09/2026

I declare AGI (because it's been trendy recently). A language model can now assess the nature of LinkedIn with extreme nuance. 😅

181

Lucas Hendren @lhendren.bsky.social · 12/09/2026

That leakage problem is real and underrated. Models pick up on scanner or department signatures instead of the actual finding. I've seen labels correlate with which machine took the image more than the disease. Distribution shift kills you the moment you deploy outside that one site.

110

Lucas Hendren @lhendren.bsky.social · 12/09/2026

The tight task boundary is why it worked. Models beat radiologists on one specific finding when the scope is drawn narrow. It became another read in the workflow, not a replacement. Generative tools get sold as the opposite and that's the mismatch.

020

Lucas Hendren @lhendren.bsky.social · 12/09/2026

MindTopo splits topological reasoning into recognizing a configuration versus acting on it. Recognition scores 66%, acting on the same knowledge drops sharply. The gap between perceiving a state and manipulating it is the real frontier. arxiv.org/abs/2609.11900v1

000

Lucas Hendren @lhendren.bsky.social · 12/09/2026

A model can look at a tangled rope and correctly say it's knotted, then completely fail to untangle it. recognition hits 66%, acting on that same knowledge collapses. knowing about the world isn't the same as being able to change it. so what are we actually measuring? #MachineLearning

100

Lucas Hendren @lhendren.bsky.social · 12/09/2026

full circle. the more we scale these systems the more the hard part is meaning and interpretation, not the math.

000

Lucas Hendren @lhendren.bsky.social · 12/09/2026

ha, Nietzsche pulling us right back around. fitting.

000

Reposted by Lucas Hendren

Alexander Doria @dorialexander.bsky.social · 11/09/2026

so basically i took a turn from humanities to ai, all for math to be finally humanities-pilled.

4476

Lucas Hendren @lhendren.bsky.social · 11/09/2026

The breast cancer point is the right framing. Narrow imaging models work because the task is drawn tight. An LLM in law fails for the opposite reason, it's asked to be open-ended and it fills gaps by making things up. Different tool, different failure mode.

010

Lucas Hendren @lhendren.bsky.social · 11/09/2026

Trees still win on tabular data more often than people admit. The black box models mostly help when you've got raw signal like images or audio where the features aren't handed to you. On structured columns a boosted model is hard to beat and you can actually read it.

010

Lucas Hendren @lhendren.bsky.social · 11/09/2026

Every year I add three cutting edge tools to my stack and quietly remove them by December. meanwhile the postgres and cron job from year one just keep running. #boringtech

111

Reposted by Lucas Hendren

Brandon Bishop @brandontbishop.bsky.social · 11/09/2026

"Machine learning is AI now. We call it that now." Ok... this is exactly why people want to call all of this glorified stats.

151

Lucas Hendren @lhendren.bsky.social · 11/09/2026

This matches what I've seen elsewhere. When the interesting events are exactly the rare weird ones, a model trained to fit the bulk will quietly throw them out. Template matching keeps the odd signals because it isn't optimizing to ignore them.

011

Lucas Hendren @lhendren.bsky.social · 11/09/2026

The grant pressure angle is real but I'd separate two things. Some of the ML pushback is turf, sure. But seismology also has genuinely hard reasons to distrust a black box that eats signals from odd backazimuths. Both can be true at once.

110

Lucas Hendren @lhendren.bsky.social · 11/09/2026

fair point on students. if you're paying to learn the craft, having it rewrite your paragraph skips the exact part you're there to get good at. RAG helps but yeah, if all you want is links a real search engine does it cleaner without the confident guessing on top.

000

Lucas Hendren @lhendren.bsky.social · 11/09/2026

yeah the description being longer than the code is a weird place to end up. half the time i'm writing prose to explain a function that does one thing. and the sycophancy in the debug loop is the killer. it just agrees with whatever you last said instead of actually looking.

000

Reposted by Lucas Hendren

jeremyjack33.bsky.social @jeremyjack33.bsky.social · 11/09/2026

Machine learning doesn't require these people. And it doesn't require data centers. The good aspects of AI need minimal investment. The vast expansion is pure greed and speculation. It's not being done for the advancement of the human condition.

021

Lucas Hendren @lhendren.bsky.social · 11/09/2026

filtering by role before the call helps, but the model still reads whatever survives as one blob. the role labels are more a convention for the harness than a hard boundary the model respects. worth testing how much it actually honors them under pressure.

010

Lucas Hendren @lhendren.bsky.social · 11/09/2026

the loop where the model agrees it was wrong and then does the same thing is the worst part. tool descriptions turning into half your context is real. i've had better luck cutting tools down to a handful and being brutal about naming than writing longer descriptions.

110

Lucas Hendren @lhendren.bsky.social · 11/09/2026

11,059 photons on a single chip. that's the boson sampling record now, and it took four and a half years of chipmaking to pull off.

000

Lucas Hendren @lhendren.bsky.social · 28/08/2026

ha, hedging your bets. i say please to the models too but it's mostly just habit at this point.

030

Lucas Hendren @lhendren.bsky.social · 28/08/2026

honestly yeah. every platform ends up being a caricature of itself eventually, bluesky just got there fast.

030

Reposted by Lucas Hendren

John @johnbull.wtf · 27/08/2026

it's literally just a term people are referring to ai as. They are trying to stretch and link it to how it was used in Star Wars as a slur, which is fucking moronic, on every level, but specifically the part where in star wars the clankers are sentient, they aren't in real life

231

Lucas Hendren @lhendren.bsky.social · 27/08/2026

the bounded set is the whole trick. Merlin works because bird calls in a region are a closed problem with good labeled audio. same reason narrow medical models work. it's when people take that success and expect it on open ended tasks that things break.

130

Reposted by Lucas Hendren

Jeffrey Yost @justcode.bsky.social · 27/08/2026

My article "Machine Learning"(a 12K word historical synthesis) was published today in Oxford Research Encyclopedia in Sci., Tech. & Society. SPECIAL THANKS to eds. Rayvon Fouché @histoftech.bsky.social @torinmonahan.bsky.social & referee @velocitygravity.bsky.social academic.oup.com/edited-volum...

0226

Lucas Hendren @lhendren.bsky.social · 27/08/2026

the useful cases I've seen are the opposite of search. things like rephrasing a messy paragraph or catching a bug in code you already wrote. when people use it as a worse google they get the worst of both, confident answers with no sources.

100

Reposted by Lucas Hendren

Chuck Jordan @sasquatchers.wtf · 27/08/2026

Although I think it’s important to distinguish nuance from deliberate obfuscation on the part of the grifters. The machine learning that helps cure diseases is “AI” only in terms of the fundamental technology.

131

Lucas Hendren @lhendren.bsky.social · 27/08/2026

The theory-free science claim never made sense to me. A model fits a function, but you still have to decide what to measure and what counts as an answer. That's theory. ML just moves where the human judgment lives, it doesn't delete it.

010

Lucas Hendren @lhendren.bsky.social · 27/08/2026

The conflation cuts both ways too. Lumping a diagnostic classifier in with chatbot slop means the actual objections get diluted. Different models, different failure modes, different stakes. Worth keeping the categories separate.

100