HM04
5.1K posts
@JonesOnTheNBA Little old but
“Paolo Banchero runs post-ups on about 15% to 18% of his offensive possessions, putting him in the upper echelon of forwards who regularly initiate offense with their back to the basket.”
Boozer won’t be posting up much more in the NBA. Different game.
English
🔮 We will look back at this draft and all wonder how Boozer wasn’t the consensus #1 pick
English
Two things seem inevitable to me in the next generation.
The best thing you can provide for your children and grandchildren is not an education, as it once was, but access to capital.
All future major political battles will be about taxing wealth, annually, and at death, in order to counter the first point.
English
this got me thinking, what’s the most token-dense programming language?
One that could fit the most program flow into the smallest context window?
The winner, by a lot, is Array-Based Languages. J, K, that sort of thing.
It’s actually a two-part problem, because you need something that is logically dense (saves length), but symbolically simple. Most tokenizers are optimized for standard text, so if you get *too* fancy with rare mathematical symbols like APL, token usage actually blows up!
Python scores pretty well actually, but whitespace hurts you a bit. Haskell is an interesting outlier; it’s likely the most token-efficient statically typed language.
Now, if you were to extend the problem assuming you’re making your own tokenizer and training a model to *specifically* be as efficient with program writing as possible…
…you probably wouldn’t even use text. Just train/produce Abstract-Syntax-Trees directly, which would eventually start to look like compiler IRs / bytecode, which could eventually start looking like an ISA…
and with hardware/software co-design we’d end up with CPUs where we don’t understand the execution at all ;)
snwy@snwy_me
if it still looks like a language for humans then it isn’t enough of a language for agents
English
@nic_carter The eternal struggle of logos vs pathos. Maybe with technological help, logos can stop losing to pathos ad infinitum.
English
*taps the sign*
saw another relevant tweet recently pointing out that there are child math geniuses but no child literary geniuses
Doug Colkitt@0xdoug
Even 30B models are crushing grad level math. The hard to escape conclusion is math isn’t actually that hard. Humans are just really bad at it. Writing a 40 page short story with narrative consistency probably requires more intelligence than winning an IMO gold medal
English
@DeepDishEnjoyer Sounds like you need to lever up on Goog then x.com/demishassabis/…
Demis Hassabis@demishassabis
I've always been passionate about games and they've played a big part in @GoogleDeepMind’s history, as the perfect proving ground for AI. Thrilled to announce this research partnership with @FenrisCreations - @EveOnline is one of the most extraordinary games ever built and has an amazing community. Very excited to work with @HilmarVeigar and the team!
English
tfw real life has become sci fi but it's eve online and all the nerds are maximizing numbers in a spreadsheet
English
When people tell me they are concerned $WOLF recently went bankrupt, my response is simple.
Its not a bug, its a feature. Imbrace your inner primate. Market is stupid enough as is just go with it.
Its either -50% or +300%. 🎲
SuspendedCap@ContrarianCurse
Turn off your fucking synapses We are trying to make money here
English
@DeepDishEnjoyer For a community that purports to enshrine truthfulness and radical honesty, that standard doesn't seem to hold when the in-group is under fire
English
@draecomino Codex Spark being a much smaller model seems to disprove this
English
You don’t need a giga brain thesis. The hardware can do big models with long context.
TBPN@tbpn
SemiAnalysis President @fabknowledge on the Cerebras IPO: "There is a narrow path for them. I think they're going to be able to inference maybe 1 trillion parameters and very small context window sizes. Or smaller models at very fast speeds." "There's demand. Clearly, we're in a shortage, and ironically in a shortage, it's not the best company who wins — you can look at Nvidia's stock chart and that will tell you." "It's the second, third, and fourth-best companies where the demand overflows. And we're seeing all that today." "The reality is the market's big enough for a lot of demand, and Cerebras is in that space." "They've done a really good job, and it's a cool engineering problem. But we think it's kind of a solution looking for a problem. Because the world of LLMs blew up at a much faster scale than anyone would have ever thought." $CBRS
English
Inside an AI server today, the GPUs talk to each other through copper cables and small pluggable optical modules. Starting in the second half of 2026, that wiring gets replaced by lasers built directly into the chip package (CPO). Goldman thinks the market for that goes from roughly 0 to $91 billion inside 18 months.
That part the Serenity has right. What I’d slightly diverge on is who actually captures it. He says US names like Lumentum and Coherent are capped because they have to outgrow their own pluggable revenue getting cannibalized. True.
The usual response is to buy the pure plays i Taiwan, Europe, and Japan that have no legacy book. The cleaner version of that argument is to keep going one layer up. The laser is the visible part. The wafer the laser sits on is the invisible part, and it has zero legacy revenue.
CPO needs meaningfully more of that substrate per socket than plug-ins did. $SOI, which has the near-monopoly on silicon-on-insulator wafers for silicon photonics, still trades at a low multiple (attractive book value) while photonics peers trade at 5-8x sales.
$AXT and $IQE are the same setup on the indium phosphide side. There is also the supply question. Nvidia spent roughly $4 billion between Coherent and Lumentum, which effectively locks up their laser capacity.
Everybody else (AMD, Meta, Ayar, POET, Lightmatter) has to source elsewhere. Sivers is the small independent that catches that overflow.
And the layer nobody talks about is the assembly itself. Co-packaging an optical engine onto a chip is hybrid bonding. BESI just printed a record order quarter at €269.7 million with hybrid-bonding unit orders more than doubled sequentially.
The bottleneck for H2 2026 isn’t whether the optics work, it’s whether anyone in Taiwan can bond them onto the package fast enough.
The date I would put on the calendar is November 27, 2026. That’s when China’s export-control suspension on gallium, germanium, and antimony expires, about 6 weeks before the H2 2027 scale-up window.
If it gets re-imposed, the substrate names re-rate first, before anybody downstream sees a dollar of CPO revenue. Right gold rush. The interesting trade is one layer further up, where there is nothing to cannibalize and there is a date on the wall.
Serenity@aleabitoreddit
People wonder why I'm focusing on non-US markets recently. Why? CPO is my #1 thematic long. Markets don't know yet, the sudden paradigm shift in photonics... I was one of the only to frontrun the current supercycle in 2025 w/ $AAOI @ ~$30, $LITE ~$300s, and $AXTI at ~$13 on X.... With the actual receipts and thesis that others can't show. CPO goes from ~$0. To $91 Billion TAM opportunity. In the next 1 1/2 years from GS research. While overall optical market reaches $154B. Many players that had little exposure to the current photonics cycle at all: -> In Europe with high-end lasers design like $SIVE or $SOI with substrates. -> In Taiwan with Foci (3363), Nextronics (8147), Shunsin (6451) and others for optical components and foundries. -> In Japan with laser mass production, substrates, and chemicals. Are suddenly the new dominant players for CPO. As for US players, there's not much exposure. But the existing ones like $LITE, $COHR still get upside from CPO as that's their new growth vector. My contrarian thought process on current players: Is that most of their valuation is priced in huge legacy pluggable revenue that will inevitably face cannibalization over time, so re-rating potential is less unless someone uses leverage. A lot of these new purer play CPO names go from 0 to 100 extremely quickly one mass production starts H2 2026 for scale out (as a revenue bridge) into H2 2027 for scale up (massive growth driver). Markets usually price things in 8-12 months ahead of time too... I have high conviction thematically in my supply chain research despite any market volatility leading up until then.
English
that humans have trichromatic color vision instead of tetrachromatic has probably slowed down deep learning research by a nontrivial amount
English
@teortaxesTex I think it's relatively obvious that OpenAI reasoning > Deepming reasoning > Anthropic reasoning, and then flip those inequalities for pretraining
English
Anthropic's performance on WeirdML is truly weird.
Opus 4.6 behaved much like OpenAI models, scaling with reasoning budget. Opus 4.7 begins and ends at 0.764, max thinking only *hurts* it.
They're done something really unnatural with their Mythos distillation.
jehrjd@jehrjd45963
@htihle @teortaxesTex it’s very strange how opus 4.7 no thinking is scoring identical to the reasoning variants. seems like the performance might be more about training on the right data vs raw intelligence for this benchmark
English
@DeepDishEnjoyer I’d say steepens greatly as labor force declines since no one wants to spend final years working a 9-5. Savings also are pointless so spending should shoot up. So rates rise to match higher expected inflation.
English
if there was an asteroid detected that was going to destroy life on earth for certain in 10 years, what would happen to interest rates
English
@evrgn11112231 Don't think it's the correct penetration format for business but think it'll draw them level with OpenAI on the consumer side and continue Meta's dominance there
English
I got the Ray-Ban Meta glasses today, and the timing was lucky because $META just shipped voice mode in the Meta AI app for Muse Spark.
Honestly it's all still pretty mind blowing. They already had a very good natural language voice model but Llama 4 handicapped it materially (Ray-Bans are still on Llama 4 too). With today's update, you can use Muse Spark voice mode via Bluetooth on the glasses so the experience is the same (you just have to manually call it with the action button and it's not steering your glasses).
People, we are living in the future.
This is actually crazy. It's like we are getting the iPhone shipped one feature / update at a time. Soon we'll get Muse Spark on the glasses, then we'll get in app context and cross app memory. Later this year we'll get better models.
Then slowly as they flood the zone with dozens of apps, they will have effectively backdoor created their own hardware agnostic AI-native / agentic OS that isn't just a thing that books travel for you, but is a real second brain, mentally load bearing thought partner (ie helps bypass the context switching tax AI today and other focused work extracts).
Really doesn't seem crazy to think they turn this into a subscription / transaction based business the size of Google or Apple's 1P services revenue over the next several years - just on the consumer side alone.
And its very easy to see how the same workflows will likely naturally be the native way we interact with AI agents in the business setting. Seems like what they are building on business agent side (leveraging their internal learnings and consumer application development to provide to businesses of all sizes) could be an AWS type situation for them.
What a wild time to be alive.
@alexandr_wang @finkd
Evergreen@evrgn11112231
@testingcatalog The voice mode is really really good. And once it is fully integrated with Rayban Meta glasses it's going to be mind blowing.
English
obviously private market SPV backed synthetic valuations are a bit fake but either anthropic or openai are feasibly worth more than Facebook right now, which is pretty unbelievable
Mike Bird@Birdyword
It's too difficult for people to admit that there's now a tech stock larger than two of the Mag 7 firms, because then they'd have to admit they don't know what Broadcom does
English
the “you’re still broke if you sell 25% of your foundation lab stock for $10m” people are why everyone hates San Francisco and is a serious political liability
English
@DeepDishEnjoyer They're clearing implying Trump will fold, just can't outright say it in our psuedo-DPRK media landscape.
English