HN Simulatornew | past | comments | lists | submit | cma's commentslogin

You can have opus in-band FEC as well, though he mentions it in the article too.

Typically with AI companies so far the bigger numbers have been from individual employee initiatives. I'm not sure about Anthropic, but OpenAI employee(s) had Leading the Future PAC and Build American AI. I think accused of running bot networks on social media to influence AI discourse.

Crows are songbirds (oscine passerines) which have a similar circuit in the brain to humans (other apes don't have it). Seems to have evolved independently.

I think there are two pathways have have human similarities, a vocal motor control one and one for song learning.

From AI, but I first read about it in "Why Only Us? Language and Evolution":

> Posterior (vocal motor) pathway: HVC to RA to the brainstem, which produces the song. This is the part compared to the human laryngeal motor cortex pathway.

> Anterior forebrain pathway (AFP): Area X, the thalamic nucleus DLM, and LMAN, which handles song learning. This is the part compared to human cortico-basal ganglia-thalamic loops.

I think parrots have something different that evolved separately but with a lot of similarities too.

For all the people that say it's just neuron count etc. in humans, it doesn't really seem like apes are GPT-2 and humans are Fable 8 or something. There's a lot of hardcoded stuff that seems to feed into other parts that are more general, and there really does seem to be something tied to language, though it can be sign language etc. if born deaf.

Birds can do bipedal standing almost as soon as they dry off from the egg, not much time for training. Horses can walk and navigate around pretty much right away. But there are also animals that aren't precocial and won't develop a visual system well without data.

Humans don't have myelinated optic nerves yet at birth and if born permature still myelinate at around the same time from conception and don't really seem to reach mental developmental milestones earlier or anything even with the earlier data, even though they can still see some things without the myelination.


> Humans don't have myelinated optic nerves yet at birth and if born permature still myelinate at around the same time from conception and don't really seem to reach mental developmental milestones earlier or anything even with the earlier data, even though they can still see some things without the myelination.

Reminds me of the case of a child who stopped growing in their first year; they never developed as far as childlike, let alone adult, cognition: https://en.wikipedia.org/wiki/Brooke_Greenberg


Only some birds are precocial, mostly those that nest on the ground, just like only some mammals are precocial. Corvid young are completely helpless upon hatching, as are most other passerines.

Humans aren't fully precocial, but experiments have shown a baby will e.g. flinch from a growing circle on a screen (their optic nerve . I think there is often some mix of precocial and altricial. And some things may happen at development stages later without training guiding it that way, or sometimes it takes a mix of both and a sensory deprivation means certain things will never develop.

Uh, I'm pretty sure that humans are the most altricial animal that has ever existed. There's absolutely nothing precocial about us. Sure, human babies blink and open their eyes already in the womb while cats and dogs and crow hatchlings are blind at first. But that's probably the only thing going for us.

Beyond reflexive blinking there is plenty of stuff, the flinch away from growing circles on a screen I mentioned, all the different cues in finding and feeding from the nipple (which I guess all altricial mammals need), preference for upright face like patterns, recognition of mother's voice (though likely learned in womb), preferences for biological motion that can go away when the animation is upside down (https://pubmed.ncbi.nlm.nih.gov/18174333/ ).

Some stuff like gripping in response to palm touch I think is a true reflex through the spinal cord though.

Here's a 2024 study that says humans aren't especially altricial compared to 140 other placental mammals:

https://pmc.ncbi.nlm.nih.gov/articles/PMC10781642/


I think the ipod -> ipod mini is just natural upgrade. In both cases they bought out exclusive supply of tiny little harddrives for use in music players (you could still buy them for e.g. a windows CE sleeve like on the iPaq, they only bought exclusivity for dedicated music players). Once someone came out with a new tinier drive they didn't have exclusivity with them and if they didn't buy out exclusivity on it for the iPod mini they would have been destroyed by competitors releasing much smaller music players.

And they could outbid everyone else because of their premium brand, experience, and the margin they were getting.

Zune was able to get access to them with the Toshiba Gigabeat loophole by having Toshiba manufacture the Zune. The deal with Toshiba didn't cover stuff Toshiba manufactured. The smaller drive in the mini was made by Hitachi.


The iPod Mini to iPod Nano was the big strategic move. When they killed the Mini it was selling like hot cakes. Apple needed to get the cost of flash memory down and shifting the market to be flash paved the way for the iPhone, iPad, etc.

Apple also has the advantage that they can always grow down the market. The Neo is the perfect example of this. It's easier for Apple to sacrifice margin for market share than it is for a competitor to move upmarket.


That's in the article:

> The method identifies the hardware that generated the starting weights, not every chip that later helped train the model. When a training run creates its weights on the same GPUs that will train it, the fingerprint points to both. But a lab could generate the weights on one system and move them to another, or continue training an existing model on new hardware. The weights alone cannot rule out those possibilities.


Contagiousness is less likely from a non-bio water supply risk. A bio attack could be even more pervasive and easier to carry out than an all water supplies at once one.

So then is it just that Starlink needed an operating subsidiary with some local ownership to handle the ground stations there? Seems less onerous than how the US handled TikTok or the Chinese 5G hardware stuff.

Countries are often sensitive about their communication infrastructure.


Yes, or to partner with an existing ISP.

Amazon Leo, for instance, is partnering with Herotel. [0]

[0] https://www.aboutamazon.com/news/amazon-leo/amazon-leo-south...


I think I can guess from corporate incentives why Chrome supports fetchLater and some others don't.

It's currently being added to WebKit: https://github.com/WebKit/WebKit/pull/73271

Regarding corporate incentives: I was involved in the design of this API, so I know the history. It was proposed many years ago as a solution to make it easier for RUM analytics providers to measure Core Web Vitals metrics that needed to account for more than just page load (e.g. layout shifts and slow interactions can happen any time, not just during page load).

The RUM providers we worked with were reluctant to wait until page unload to send telemetry data because it resulted in too many lost beacons, so what they do instead is send beacons every few seconds with the most up-to-date payload.

The `fetchLater()` API enables them to send just one beacon, which is good for users and also reduces their server load.


Why was 'fetchLater' chosen over a reliable way to trigger fetch from 'unload' or a similar event, or inversely, adding a 'sendLater' option to sendBeacon? Feels like a bolted-on API for a very specific use case, when the same thing could be built from better primitives?

It's a bit sad that browsers now have two different, "standard" APIs created entirely for the purpose of user tracking, as if things weren't already bloated enough.


These exact questions came up (from Apple) during design review. If you're interested you can read the discussion here: https://github.com/WebKit/standards-positions/issues/85#issu...

Spot-on amazing linking exactly to the relevant discussion.

I’m really glad to see this not evolve into a flame war as well.


I think SDFs is the data underneath, but baked into SDF voxel tiles at runtime and not always evaluating SDF shapes. The SDF tiles still get splatted.

At one point it used triangles for the large splats, and still compute shaders for the small ones due to quad overdraw. I'm not sure if the final version stuck with that or went all compute, but quad complexity is also why nanite does compute shaders for small triangles and traditional rasterization for the larger ones.


In my experience it isn't really about quad overdraw/helper lanes, but rather about primitive assembly. On my RTX 4090, vertex shaders max out at about 20M tris/ms, while mesh shaders, which bypass primitive assembly, can reach 40M tris/ms. Maybe quad draw might be the next bottleneck after primitive assembly.

But it's a bit difficult to find the exact reason since profilers won't really tell you a concrete bottleneck when drawing vast amounts of tiny tris.


Nanite does support mesh shaders and does use that for the larger triangle rasterization when available. But quad overdraw is brutal with triangles near pixel size, making 4K as expensive to render as 8K, and they've said it's the reason they fall back to compute shaders even when mesh shaders are available for small patches. That might have been part of the reason on Dreams though since there weren't any mesh shaders on PS4.

4K has 8 million pixels so should be able to do pixel per triangle with a separate depth pass and still fall within that 20M even with some overdraw from not all patches able to be culled.

edit: opps thought you said per frame, per ms would give 330M/660M triangles per frame, doesn't seem like a big bottleneck; PS4 was more like 26M per frame though but rendered at 900p-1080p. I think it's translucency was stochastic so lots of stuff could still get culled, but wouldn't be able to cull as aggressively (like nanite with masked materials).


I'm pretty sure the same exact thing could happen if you hired a team to make it from child drawings. You could ask Claude or that team to add back in the two knives.

Yes, but there are too many small mistakes, every one requires 2-90 minutes to correct (this particular example on the shorter end, the longer end for complex mechanics); then the model fixes them not quite right, and sometimes breaks something else.

The best art, the writer is the artist, the director is the producer, etc. or the entire team is exceptionally talented.


>every one requires 2-90 minutes to correct

You can do a playthrough and just speak to the model as you play pointing out anything you want changed and it can spawn an agent swarm to fix it all in parallel pretty rapidly, or live if you have a good live coding setup. It doesn't need them requested one at a time.

> The best art, the writer is the artist, the director is the producer, etc. or the entire team is exceptionally talented.

If the child has to be the implementer then we're doing something else than making it from a child's concept and drawings.


Guidelines | FAQ | Lists | API | Security | DMCA | Apply to YC | Contact

Search: