HackerTrans
TopNewTrendsCommentsPastAskShowJobs

thrtythreeforty

3,044 karmajoined hace 7 años
Embedded systems engineer - me [at] thirtythreeforty.net

Submissions

Designing AI Chip Software and Hardware

old.reddit.com
3 points·by thrtythreeforty·hace 4 meses·0 comments

comments

thrtythreeforty
·hace 4 días·discuss
I don't know JAX, but can this trick be applied as a higher-level function to autodiff-capable languages like JAX?
thrtythreeforty
·hace 4 días·discuss
Surely SIMD combined with multiple streams would beat both approaches. (This would be separate streams in each SIMD lane and separate streams in different SIMD variables.) There are multiple SIMD execution units, just like the 6 scalar units you mention. The latency of SIMD ops will be similar to scalar, except in cases you mention like shifts.
thrtythreeforty
·hace 9 días·discuss
Cyclists doing whatever they feel like is status quo everywhere. Now I'm most important sidewalk user! Now I'm a car! Now the traffic rules don't apply to me!

(Not all cyclists do this. But the rude ones are common enough that "cyclists" have gotten this reputation.)
thrtythreeforty
·hace 18 días·discuss
For reticle-limit chips, it's on the order of 100. And less than that once you filter out bad dies.
thrtythreeforty
·hace 2 meses·discuss
Reports indicate that cheaters need a full Windows reinstall to recover?

> installs malware

> malware does malware stuff

surprised_pikachu.jpg

Of course it's hard to write sympathetically about cheaters, but this is exactly the sort of nonsense you'd expect from installing a hostile kernel extension.
thrtythreeforty
·hace 2 meses·discuss
Can't believe I had to scroll for this. Trust the OEMs to do something very close to what enthusiasts might want, then immediately torpedo it with "but we won't sell it to you."

This level of conversion isn't exactly trivial but it also isn't rocket surgery for the kind of person who pulls an engine out for rebuild on a classic car project.
thrtythreeforty
·hace 2 meses·discuss
It kinda does have that effect, because large companies are allergic to AGPL by policy, even if they otherwise have processes to get GPL software into use. The very companies that have the pocketbook to pay for your software, are also structurally incapable of using it as FOSS. Smaller ones that are more agile about how they incorporate it, have less willingness to pay for a different license.
thrtythreeforty
·hace 2 meses·discuss
Here's a question: I don't understand the gap between these LLM powered voice agents vs CLI coding agents, the latter of which are obviously useful and quite resourceful at getting something done when asked in plain English.

Seems like an agent given 20-30 tool calls like "read_sms" "matter_command", and "send_email" would be able to work out what to do for things like "set the house to 72° and text Laura that I did it."
thrtythreeforty
·hace 2 meses·discuss
This is a weird comment because I feel the same about getting macOS to a useable place.

I probably have 5 or 6 things installed on my Mac like Scroll Reverse and Rectangle, just trying to beat the window manager into something that resembles useable.
thrtythreeforty
·hace 3 meses·discuss
I noticed this almost immediately when attempting to switch to Opus 4.6. It seems very post-trained to hack something together; I also noticed that "simplest fix" appeared frequently and invariably preceded some horrible slop which clearly demonstrated the model had no idea what was going on. The link suggests this is due to lack of research.

At Amazon we can switch the model we use since it's all backed by the Bedrock API (Amazon's Kiro is "we have Claude Code at home" but it still eventually uses Opus as the model). I suppose this means the issue isn't confined to just Claude Code. I switched back to Opus 4.5 but I guess that won't be served forever.
thrtythreeforty
·hace 3 meses·discuss
Then they get to keep both pieces!
thrtythreeforty
·hace 4 meses·discuss
This is a fantastic explanation. Thank you. The only part I am not following is how it is guaranteed that 1 bit is sufficient for the error value. Is this something the Lloyd-Max algorithm is responsible for ensuring? (Seems to me that if your quantization algorithm is crappy enough, you could need a large number of bits to store the error.)
thrtythreeforty
·hace 4 meses·discuss
I would love to lobby to change how the law works for these cases: for some definition of "firmware" (informally "software that ships with hardware and is not intended to be selected by the consumer like a computer operating system"), add a copyright exception so that modifying the firmware in situ is treated like modifying the physical hardware, because in practice they are in fact the same thing: a single component that does a single thing.

With this, the John Deere approach to gatekeeping vehicle repair would no longer be legally protected by the DMCA or by copyright law. All the other protections afforded by copyright law would still apply: you cannot rip the firmware off the hardware and distribute it, the manufacturer is under no obligation to help you modify it, etc.

However, tools which patch or circumvent antifeatures of the firmware would now be legal to use on hardware you own: it would be legal to patch out software locks, retune engine computers, etc.
thrtythreeforty
·hace 4 meses·discuss
Ooh this gives me an interesting passive-aggressive idea to counter pointless "is this still relevant" questions. "No, I haven't hit this in the last 2 days." "No, I haven't hit this since I gave up trying to do it with your tool." And so forth.

The less passive-aggressive version is to use this obviously-unhelpful answer of the obviously-unhelpful question, to actually have a conversation to get the PM to recognize that the default state of a ticket is in fact "no change." Ultimately that may turn into a stale bot if the PM realizes the policy they actually want is some sort of timeout, but at least it's not a time consuming meeting!

(Note, a cathartic thought experiment, but not really good manners to actually do!)
thrtythreeforty
·hace 4 meses·discuss
There are lots of subtle LLM-ism sentence structures that you can perceive, sort of subconsciously. Spend any amount of time with Claude and you'll unconsciously start to copy them.

This essay has several: Sentence structures, heading phrasing, and yes, dashes. I expect people know to start steering LLMs away from em dashes at this point though.
thrtythreeforty
·hace 4 meses·discuss
Pay your $25, upload an APK, and it's launched. What's left?
thrtythreeforty
·hace 4 meses·discuss
The one that took us the longest was (3). He knew he needed to go, knew all the steps, and would pee in his pants just to spite his parents.

It's really eye opening and frustrating to see children be stubborn just for the sake of it. They're literally still developing and it's normal, but they are just too young to reason with. You have to create the environment where positive feedback happens, and... wait a very long time for them to work it out.
thrtythreeforty
·hace 4 meses·discuss
You're not mentally modeling the child correctly. The child has to have several things going to successfully practice using the bathroom:

1. Seeing a problem with having a dirty diaper

2. Recognizing (at least subconsciously) that on balance, using the bathroom will be easier than waiting to have your diaper changed

3. Being willing to react positively (not obstinately) to parents' reminders to use it

4. Being able to focus enough to, say, not play in the water, and old enough to practice all the steps.

And that's just to practice. Even if they're all of the above, they'll still take time.

Source: watching my now-3 year old overcome each stage listed, one at a time
thrtythreeforty
·hace 4 meses·discuss
I've got a T480s that has 40GB of RAM (32GB SODIMM + 8GiB soldered). Works fine. Is the T480 different?
thrtythreeforty
·hace 4 meses·discuss
You accidentally raise an interesting point: good, thorough public documentation, once considered a great selling point for your system, now invites automated reimplementation by competition. It would be a shame to see public docs vanish because it turns out they are literally machine readable specs.