Mark Young
412 posts
Mark Young
@vizmo
Designer, developer, scientist
San Francisco Katılım Mayıs 2008
116 Takip Edilen76 Takipçiler
Patrick is a 34-year-old orangutan at the Metro Richmond Zoo who recently became an online sensation after a video of him tying a perfect double knot went viral.
To celebrate his birthday, the zoo gifted him a royal cloak, which he tied neatly on his own 🦧
English
@jjainschigg @shiraeis That makes some sense. There's additional cues and context that are shared, but again, not necessarily processed the same way. I believe there are systematic differences between people such that there's typically a significant difference between world models
English
@vizmo @shiraeis And as she later says, this is because there’s an actual shared reality and thus broad convergence on what amounts to a shared world model. x.com/shiraeis/statu…
shira@shiraeis
otoh, and this is the part i find hopeful: the post qt’d is about the local problem: language is a brutally narrow channel. 2 people can hear the same words, yet run totally different reconstructions, and walk away feeling either perfectly clear or completely alien to each other. meaning is inferred, not transmitted whole. but there may also be a global convergence pressure. the platonic representation hypothesis is basically the idea that when very different systems get good enough at modeling the same world, their internal representations start becoming more alike. larger models don’t just perform better, but across architectures, and even across language vs vision, they begin organizing reality in more similar ways. some fMRI work suggests the strongest overlap is often not at the final “answer” layer, but in the middle of the stack, where the system is still actively building a general-purpose world model of what’s going on. in a 2025 review of 25 fMRI studies, this intermediate-layer pattern shows up repeatedly, and the broader alignment story tends to get stronger as models scale in size and capability. that’s how I reconcile this with the post qt’d: a single conversation may be nowhere near enough to make 2 differently trained systems converge. shared reality may exert a long-run pull toward similar abstractions, but the channel between 2 people is still narrow, lossy, and path-dependent. yes, sometimes you really do hit the compression limit of a relationship, but that isn’t proof that shared understanding is fake, just that it is expensive. paper links: arxiv.org/abs/2405.07987 phillipi.github.io/prh/ arxiv.org/abs/2510.17833
English
had a series of disagreements with someone i love recently where it became obvious we're running such different world models that no amount of talking was going to produce convergence. naturally i went looking for what neuroscience says about why two people can speak the same language fluently and still completely fail to land in the same internal state
the literature increasingly suggests language is way less "transmission" than people intuit. rather than sending a fully specified packet of meaning, the speaker emits a sparse, lossy signal, which the listener then reconstructs from their own priors, context, and internal model of the world.
this frame fits the results in the quoted tweet unusually well. Zada et al. show that during real conversation, linguistic content briefly occupies a shared representational space across brains, but VERY briefly. Goldstein et al. show that within a single brain, comprehension unfolds over time in a layered hierarchy that looks a lot like moving through depth in a transformer. together the story is "words are small cues that help coordinate much larger contextual states across time,"
which means the feeling of being transparent benefits the speaker more than the listener because two people can use nearly identical language, feel totally understood by themselves, and produce completely different reconstructions on the other side.
it's honestly a small miracle that two differently trained systems ever converge as closely as they do through such a brutally narrow channel.
what gives me hope is that learning how language actually functions across humans and machines might teach us something real and meaningful about improving human to human communication as well, by making the lossy channel a little less lossy.
that said, some people don’t really want a wider channel. some people just want confirmation that their reconstruction is the only valid one, and maybe the most honest response to that isn’t frustration but recognizing you’ve hit the compression limit of the relationship
shira@shiraeis
Found 2 papers on language, brains, and LLMs that together tell a story no one has cleanly articulated. One looks at spoken conversation and finds that contextual LLM embeddings can track linguistic content as it moves from one brain to another, word by word. The relevant representation shows up in the speaker before the word is said, then shows up again in the listener after the word is heard. The other looks within a single brain and finds that the timeline of verbal comprehension lines up with the layer hierarchy of LLMs: earlier layers match earlier neural responses, deeper layers match later ones, especially in higher-order language regions. Both papers are from the same group at Princeton. Quick summary of each, then what I think they mean together. Zada et al. (Neuron 2024) recorded ECoG from pairs of epilepsy patients having spontaneous face-to-face conversations. They aligned neural activity to a shared LLM embedding space and found that contextual embeddings captured brain-to-brain coupling better than syntax trees, articulatory features, or non-contextual vectors. The embedding space works like a shared codec. Speaker encodes into it before they open their mouth, listener decodes after. Goldstein, Ham, Schain et al. (Nat Comms 2025) pulled embeddings from every layer of GPT-2 XL and Llama 2 while people listened to a 30-minute podcast. In Broca’s area, correlation between layer index and peak neural lag hits r = 0.85. As you move up the ventral stream, the temporal receptive window stretches from basically nothing in auditory cortex to a ~500ms spread between shallow and deep layer peaks in the temporal pole. The classical phonemes → morphemes → syntax → semantics pipeline doesn’t recover this temporal structure. The learned representations do. Together, these papers make conversation look a lot like two brains running closely related forward passes, with speech acting as a brutally lossy bottleneck between them. Inside a single brain, the structure of that forward pass (shallow layers tracking fast local features, deeper layers integrating slower contextual information) looks a lot like the way comprehension actually unfolds over time. What's crazy is these models were only trained on text, and yet their layer hierarchy STILL mirrors the temporal dynamics of spoken-language processing, so whatever structure they picked up is probably not just a quirk of modality. It actually seems to fall out of language statistics themselves, which is not what the classical picture would predict at all. If comprehension were really a tidy pipeline of discrete symbolic modules, you’d likely expect to see that cleanly in the neural timing, but you don’t. If you take compression seriously, this suggests language is not really about explicit symbolic manipulation, but more accurately about lossy compression over a learned continuous space. Brains and transformers may be landing on similar solutions because the statistical structure of meaning constrains the geometry hard enough that very different objective functions (natural selection vs next token prediction) still push you into roughly the same region. Something I find kinda funny is transformers compute all layers for a token in one feedforward pass, while brains seem to realize something like the same hierarchy sequentially in time, sometimes within the same cortical region. Broca’s area obviously does not have 48 anatomical layers, but its temporal dynamics behave almost as if it does, which is quietly a point in favor of recurrence. What transformers learned may be right even if the brain implements it more like an RNN unrolling over a few hundred milliseconds. The field ditched RNNs for engineering reasons. The brain, apparently, did not get the memo. The better frame than “LLMs think like brains” is representing meaning in context may just be a problem with fewer good solutions than we assumed. If you optimize hard enough on language statistics, you may end up in a solution family that overlaps miraculously well with what evolution found. There’s a real isomorphism in the problem, even if not necessarily in the machinery. Paper links: pubmed.ncbi.nlm.nih.gov/39096896/ nature.com/articles/s4146…
English
@jjainschigg @shiraeis But people? No way. Not to mention that whatever the human analog of "internal geometry" is, it's not a data structure and it's got wild I/O
English
@shiraeis But I’m sure you saw the research that says that the internal geometry of differently trained and quantized systems was broadly the same no?
English
@JebraFaushay His singing is still great. The show's production values look awful
English
It’s ok to retire. You don’t have to perform anymore.
Here is Rod Stewart strutting around with his tight little flare jeans and teased hair.
English
This is going to happen like gravity, with comical and terrifying results. For a brief, shining moment, some folks will think “omg, everything talks to everything! Everything is self-healing! Look at my alarm system warning of a fire by Van Eck phreaking a connection to my refrigerator!” And then the horrors will begin.
lachlan@hyprturing
i think chips with burnt-in LLMs that run at a very low power will probably result in much of the world around us being unneccesarily intelligent. cheaper to throw that chip and some flash with a readme into an automatic door opener than develop firmware for it.
English
@TechLayoffLover I doubt Meta is both this clever and this dumb at the same time
English
Meta just confirmed a 20 percent company wide headcount cut but insiders are screaming the real story is way uglier than the efficiency memo.
Direct word from multiple sources deep in the org. The announced 20 percent is only the near-term floor.
A senior manager told me it's the starter number. Top leadership is already modeling two to three times that long term.
They dug into developer activity logs after rolling out heavy AI tools across engineering. What they found floored them. Hours worked have plummeted for a huge chunk of the org. Many seniors now clocking under 10 hours of real productive time per week because the tools are handling so much grunt work so fast.
One staff engineer said leadership pulled dashboards showing commit rates holding steady or climbing while logged hours tanked. Same output. Way less human sweat. The math was brutal. If AI is letting one person do the work of three with half the time input why pay for three.
They have been quietly running knowledge extraction sprints for months. Senior engineers screens recorded 24 7. Every prompt logged. Debugging flows captured in brutal detail. Entire decision trees filmed under the banner of process documentation for continuity.
One engineer described being forced to whiteboard his whole system design playbook. Trade offs. Failure modes. Scaling hacks. All while cameras rolled. They called it knowledge transfer to support the transition.
The transition is agents and remaining skeleton crew armed with those exact recordings prompt libraries and heavy Claude+Gemini access. Replacements are already shipping changes 40 percent faster using the precise workflows they ripped from the outgoing engineers.
They are not slashing for cost alone. They are turning 15 plus years of Meta engineering DNA into structured training data. Every document your process request is not teamwork. It is feeding the beast that replaces you.
The strategic AI pivot line they are feeding the press is cover. They are replacing the entire engineering organism with agents trained on their own seniors captured minds.
CTO level playbook is already locked. Extract. Document. Automate. Repeat at two to three times scale.
If you are still at Meta and someone pings you to please record a quick walkthrough of your workflow for the team run. Do not document. Do not hand over the keys.
The knowledge extraction is complete.
If you are inside watching more engineers get gutted because AI made their 40 hour week look like 10 hours of value, DMs are wide open.
English
The tiger Maruay lived confined in a cage with a cement floor and was rescued. Since then, he is thriving and loves to relax in the lake with his red ball.
English
@jjainschigg I guess I'm doing it wrong. Agents responses vary a lot. Will I get used to that variance? Is the future sloppier?
English
@3YearLetterman If you're ever in Oklahoma is it a way to move around the law?
English
What the hell could possibly be the point of this?
Amazing Maps@amazingmap
Passenger rail networks in the United States vs Europe
English
@PSxMania_ I loved the first Jak and Daxter game so much - work of art, lots of soul, so much fun
English
It's January 1, 2026. The basic premises and much of the detail predicted by the movie Her, made in 2013, are all now real, and right in front of your face.
How's everyone feeling about that?
English
@elonmusk There's this weird numerology where some people think that people that don't agree with them are "low IQ"
English
@DeepDishEnjoyer I wouldn't be surprised if people in Boston have a wider variety and higher levels of education on the average given the abundance of universities there. But I'm fairly certain you can't measure someone's IQ in a typical conversation
English
I'm slowly coming to realize that a big reason I enjoy being in Boston so much is that the baseline IQ feels at least half a standard deviation higher than SF. It's this intangible aspect to every conversation across a wide range of personalities, like a brain fog was lifted.
English
@swardley @leashless Rich people's lives in space will suck compared to middle class life on Earth. Are they that naive?
English
The problem with this, is we knew over 15 years ago that a certain group of wealthy individuals thought their salvation lies in space (think, the film Elysium).
They're not going to fix the planet, they have no intention of fixing the planet. Never have.
Still, the launch was picture perfect. Spectacular.
Dale Vince@DaleVince
Watching obscenely rich people and their friends blast off into (not quite) space for a jolly is bizarre when there’s so much to fix here on earth. This song by Gil Scot Heron song does a good job of summing it up…
English
@SenatorBanks Truth? You knew nothing about that person or the work they did yet you insult them. What a jerk. Definitely not a leader
English
🚨 BREAKING: Elon Musk says that he thinks we could have a million people on Mars in 20 years.
English
What was the Black newspaper in your hometown? Trying to see something
English