HNHacker News
TopNewBestAskShowJobs

tliltocatl

1,072 karma · joined June 23, 2024

submissionscomments
tliltocatl··on Show HN: Three genlocked RP2350B make a console – 3k sprite pixels per line)
> Or do you still need the product as a whole to be compliant?

I'm not 100% sure, but I've been thru a round on RED (on a full-custom device) and much of it involved security and authentication within application software. RED basically boils down to "we are fucking tired of your little plastic shit running botnets, the customer are clueless, so fix it if you want to sell it".

So actual code review and running tests in front of the auditors, basically everything short of a pentest, not just RF PHY stuff. So my gut feeling is that while you can skip the stuff related to RF PHY if you use an off-the-shelf module, the rest will is still be on you.

Based on quick googling the certifiers agree with me (but again, that's not same as as a court saying that's true): https://compliancetesting.com/ce-certification-for-espressif...

tliltocatl··on Samsung is expected to more than double output of its HBM4 and HBM4E DRAM
How so? FEOL is pretty much the same, BEOL is almost the same save TSVs, the packaging tech is different and more advanced, but not exactly 1:1 comparable. Do TSVs really occupy 3x the area of DDR IO's?
tliltocatl··on Show HN: Three genlocked RP2350B make a console – 3k sprite pixels per line)
Love the idea, hate the presentation. Also, wasn't able to find the simulator link, is the simulator not publicly accessible?

Also, one friendly tip - if you are EU-based and plan to sell the hardware, drop the ESP32/wireless. Yes, that would mean no wireless pads, but it will make regulatory compliance a lot easier. Maybe replace with a breakout socket (not sure if that is allowed).

tliltocatl··on MS patents system to freeze games, inject ads – commercials earn playtime credit
I mean, good thing it's patented? To evils cancel each other, I guess.
tliltocatl··on Big Tech uses guarantees to keep $300B AI exposure off balance sheets
> What is the threshold for you?

Golf courses are stupid too, but the scale is nowhere close. The difference is that new golf courses are not financed by an too-big-to-fail infinite money glitch, so they are limited by how many people are willing to spend their money building golf courses.

> If AI CEOs had better PR, would you be less anti AI?

I'm not against AI. I'm against companies blowing up a bubble. Notice that the question "is it a bubble" is orthogonal to "is it useful".

> why you think AI can’t make a profit.

AI can make a profit - in niches. Automatic translation. Mathematic proofs - but yes math is more than proofs and there never were much money there. Maybe code generation in some cases, but I don't see "programmers becoming obsolete" any time soon - spewing out code was never the problem for me, domain knowledge was and still is. And yes, all that doesn't make it any less of a stochastic parrot. A useful parrot, but still a parrot.

But OpenAI and Antropic can't generate the profits they promise at the scale they need to break even. Because

a. Most of their use cases yet demonstrated are self-defeating. Text generation is useless once everyone hates slop and long texts stops becoming a useful signal. The whole story is basically "I can generate infinite gold, I'm now can buy anything". No, your gold is now worth very little. And everyone whose pay was indexed in gold is now fucked too.

b. They have no real moat aside from loads of money. Their entire plan is "growing too big to fail, then engage regulatory capture." And people engaging in in regulatory capture deserve nothing but slow and painful death. Not that it would work anyway, we've got China fixing that, for better or worse.

tliltocatl··on Big Tech uses guarantees to keep $300B AI exposure off balance sheets
1. Because resources (land and water) are limited and build-out crawls the resources from the more important (even vital) affairs.

2. Because there is still no convincing way how AI giants would make profit. Whatever impressive and useful tricks the stochastic parrot is able to do will never pay back enough to make a breakeven.

3. Because AI giant CEOs are a bunch of galactic-scale douchebags who keep saying "this is far too powerful for a mere mortal, but as long as we hold this power it's okay" with a straight face while presenting us with the stochastic parrot. No, seriously, I'm very impressed. Someone actively subverting the tech narrative couldn't do a better job.

tliltocatl··on Lawsuit: Anthropic, OpenAI, SpaceXAI and Google made illegal slowdown agreement
I welcome our new Chinese AI overlords.
tliltocatl··on App building platform for young kids to learn how to code
Too all good kids trying to learn how to code: stay clear of proprietary platforms. These will probably would not be there when you grow up.

And to all the AI doomers out there I've got a message, but it's not exactly in a kid-friendly language.

tliltocatl··on TSMC revealing details about next gen A14 node
> Bi-SRAM

Looks cool, but isn't it by definition planar process-only, no FinFET for you?

tliltocatl··on TSMC revealing details about next gen A14 node
> 1T-1C SRAM

Aka eDRAM? It was a thing back on 14nm node, but AFAIR not so much anymore. The 1C part is hard to get right in a way that is compatible with a logic process.

tliltocatl··on Will There Be a 7G?
> What would really help is the easy provisioning of signal boosters.

But that's also a physics thing. Plain same-frequency boosters hamper frequency reuse on neighbor cells, so less bandwidth for you. And if boosters run on a separate band, you'd need to do frequency planning for them just like it's done for cells now.

Wireless is finicky, we should really avoid using it everywhere wired can be used instead.

tliltocatl··on Show HN: What if the speed of light was 5 km/h?
We wouldn't "release" that because there is simply no way to tell if it is wrong. The past redefinitions happened because previous definitions were too ambiguous rather than "wrong".
tliltocatl··on Show HN: What if the speed of light was 5 km/h?
Fun fact: given that the meter is defined in terms of speed of light (and cesium frequency), "changing the speed of light" is a counterfactual impossibility, what is actually changing are some other fundamental constants.
tliltocatl··on Show HN: What if the speed of light was 5 km/h?
I think the drones don't use a complex encoder because latency.
tliltocatl··on Time lords prepare to kick leap seconds into the next millennium
Should have used TAI for internal timekeeping.
tliltocatl··on Hang on to Your Firefox
Screw average user.
tliltocatl··on Creepy Crawlies
It is server-rendered cgit pages, there are potentially quadrillions of unique. They are not cacheable.
tliltocatl··on Norway should buy OpenAI
Nei. Please don't. Enough money wasted already.

If you want to jump on the bandwagon just build a semi fab. Lots of clean water, lots of electricity, lots of geologically stable ground, lots of money that can't be invested locally - can't see why that haven't been done already. And it will still be useful after the fad is over.

tliltocatl··on LLM City – 3D render of all Kimi K3's weights as 2.5mm tiles
I get:

Visualization failed:

createProgram@https://magik.net/llmcity/:1098:115

MatrixRenderer@https://magik.net/llmcity/:1107:20

main@https://magik.net/llmcity/:1895:20

(IceCat)

tliltocatl··on Mistral Patent for “Code implemented tool calls”
> Patents are designed to protect costly research

I'd argue that how they should have been, but the reality is quite the opposite. The patent framework is all about novel ideas rather than costly research, even though ideas aren't any valuable anymore - the easy stuff that is cheap to reproduce once the idea is known is already discovered.

Multimedia codecs and similar stuff are pretty patent-worthy (costly research) but their value is an exponential function of their adoption, and hence the unencumbered one always wins. Same with basically everything else.

tliltocatl··on AMD acquires Taalas to boost inference performance by etching models in silicon
I think they already do that, except it's not 1980 so you don't fix the upper mask, you fix the lowest metal layer (the upper layer is very coarse and is only useful for power). But even a single mask is still quite expensive.
tliltocatl··on AMD acquires Taalas to boost inference performance by etching models in silicon
Using LLMs for information retrieval is the most stupid thing one can do. Especially when old methods work much better.
tliltocatl··on Show HN: Browser CAD for stock-tube structures with generated 3D-printed joints
Nice idea, but the interface needs some polishing:

- In general, you are exposing too much of the internal object model to the user. E. g. nodes added by the user and nodes added during reconciliation might be same class internally, but the user don't care much about the latter.

- "Reconciliation" looks like a sure way to confuse everyone (e. g. made an equilateral triangle, why are my joints not exactly the same? Forgot to reconcile one).

- A button for automatically merging all attachments at a node into a joint would be nice (but that might require keeping a track on what got reconciled where).

- Would be nice to automatically adjust length ratio and reconciled offset for shallow connection angles.

- You have to tweak visibility settings to be able to select nodes, virtual links and attachments, that's a bit ugly. Would be nice e. g. to be able to select a virtual link behind currently selected physical link, or a node out of a list of nodes on the current link. Especially when adding a connection in the middle of a link - you have to select the link, then hide physical links. Perhaps only allow selecting valid/semantically meaningful elements in this modes, and highlight them so that they aren't occluded?

- Any way to exploit symmetry? That's a bit advanced, but if we are making a regular tetrahedron it's quite a chore to repeat every operation four times every time you have to readjust it. That's particularly desirable because when you one physically manufacturing the design, one'd like to know which parts are interchangeable.

- Make automatically assigned names shorter (i. e. J1 instead of "Joint 1") so that you could put them as labels on prints.

tliltocatl··on I regret migrating to Codeberg
> Why would a developer using LLMs be hammering their own codeberg site?

Because a significant percentage of "developers using LLM" have no clue what they are doing?

tliltocatl··on I regret migrating to Codeberg
Hosting your code on a free server is not a universal right. It is a privilege.
tliltocatl··on Couple pay >$800k for a gene-editing therapy for their daughter. She died.
Neurodivergency, or whatever you might call it, sucks, at least during childhood. Ten times so in a family-centered society. I had a mild enough case to grow up functional and independent, but if not for that… I totally see how one might chose a risky treatment even when fully informed. The problem here, it looks like the consent was nowhere informed enough, and the research team chose not to publish the case after it turned out a failure.
tliltocatl··on Couple pay >$800k for a gene-editing therapy for their daughter. She died.
TL;DR:

- The girl had a really rare genetic disorder, traceable to a single-base mutation, that result in intellectual disability. Her case was quite mild, she was verbal and only had a learning disability, other cases are often far more severe.

- They went for using adeno-associated virus as vector to deliver a CRISPR payload. It caused kidney and liver damage (AFAIUK due to immune response, not the virus itself? So hard to predict from an animal model.) which turned out fatal.

- The family paid a significant share of the research funding and some off-record financial favors to the research team.

- The research team's recent publication in Nature didn't mention the case at all (they basically chose keep silent about failures).

The whole story has quite some Flowers for Algernon vibes except real life is way more cynical and sad. And I disliked the book back when I read it because it felt like a weepie just for the sake of weeping.

tliltocatl··on My husbands suicide shows theres something very wrong with US insurance industry
> with the caveat that there is NO private pay option

I don't think there is a society in the world that ever had that successfully executed in practice. I live in Norway, and even if the public heath system is probably as good as it gets here, I often use private providers for non-urgent things simply because waiting time can be up to six month otherwise.

And even if you somehow manage this on paper, you'll be very surprised on how markets work. They don't need legal recognition. They don't even need money. See: USSR. Health care was public and medics weren't even a particularly attractive profession because there were too much of them already. Still if you wanted to survive an appointment you'd better have a package of some deficit wares ready.

What denies public healthcare money isn't "wealthy trying to cut public spending down". It's bureaucrats wanting their share of the pork barrel.

tliltocatl··on ECC and DDR5
> Corolary: submarine datacenters are much better than orbital datacenters

Yea, except datacenters also need connectivity. And submarine connectivity, while not particularly expensive, is quite vulnerable - even more so than on land.

tliltocatl··on Memory Safety's Hardest Problem
This sounds pretty similar to Java's array-of-templates problem. Is it really more of a problem than parent pointers?
← PreviousPage 2 of 16Next →