Netflix does not replace television channels, except when it actually partners with TV broadcasters. Netflix only replaces how you consume series and films. Many TV programs --news, short series, specials, local/regional programs, etc-- aren't available at all on Netflix, or when they are they only become available months/years later due to licensing.
Until a few months ago, domain experts who ciuldn't code would "make do" with some sort of Microsoft Excel Spreadsheet From Hell (MESFH), an unholy beast that would usually start small and then always grow up to become a shadow ERP (at best) or even the actual ERP (at worst).
The best part, of course, is that this mostly works, most of the time, for most busineses.
Now, the same domain experts -who still cannot code- will do the exact same thing, but AI will make the spreadsheet more stable (actual data modelling), more resilient (backup infra), more powerful (connect from/to anything), more ergonomic (actual views/UI), and generally more easy to iterate upon (constructive yet adversarial approach to conflicting change requests).
We have monthly presentations at my job and the business folk are really leaning into AI. The biggest win so far are them being able to generate new user experiences and get them into figma by themselves. They're able to test a design, get it into figma, generate some code, and get it in front of users without a developer or designer at all. It's not perfect but the tests show what we need to focus on vs what falls flat when put in front of users. It's very impressive and I'm proud of them.
Super interesting. I don't know why, but something about this comment made something click for me, as an "AI fatigued" engineer.
From the view you describe, it seems AI just lets you experiment faster, when all you want to do is experiment. You find product market fit easier, you empower designers more, etc. Much easier to iterate and find easy wins from alternative designs - as long as your fundamentals work!
Only problem is that you are experimenting in public, so the massive wave of new AI generated features come to the public from everywhere at once. Hence the widespread backlash.
Not to mention, the core job function when you are experimenting is different from what defines a lot of hard technical progress: creating new technologies, or foundational work that others build on, is naturally harder and slower than building e.g. CRUD services on top of an existing stack. Deep domain expertise matters for selling, deep programming expertise matters for stability. I don't know, curious where the line will end up getting drawn.
Yeah, the examples I've seen really focus on experimentation which my employers's platform is designed around. We are constantly testing changes in design and copy and hoping that we get small incremental increases in user attention. AI is really suited for these small changes and it allows us developers to build platforms specific stuff instead of working on baby tweaks. We already had a pretty good system where astute business people could tweak HTML and CSS but now their lives are even easier and they can focus on their actual job which is increasing customer sign ups and attention
This is a really weird website, I glanced over a bunch of different articles and all read like AI slop to me.
Indeed, a detecting tool like GPT Zero is "highly confident" that 97% of this article is AI generated, while AI Detector returns "We are 100% confident that the text scanned is AI-generated".
Curious if this is an uncanny valley situation, because there aren't that many tells (dashes, etc.) in the text itself. Does it feel the same to you?
Didn’t look at it too closely, but the whole article as it stands is almost completely copy-pastable from a llm chat. Another comment pointing out that there’s some code that doesn’t do anything is another clue.
(Not saying it was, but if I’d ask the llm to create and annotate a HTML manipulation poc with code snippets, I’d get a very similar response.)
Edit: Pretty sure the account itself is only here to promote this page.
You're proving my point. Yes, 495 possibilities CAN be stored in 9 bits. But the article shows STRING '00000034' (64 bits) as an example, not the actual 9-bit binary encoding. That's exactly the problem - claiming bit-level compression while showing byte-level examples.
And if you look at article, nothing is binary encoded, they are all integer representations all the way down.
Please someone show me a BIT implementation of this - THESE ARE BITS 0 1 0 1 1 0 - It's called BINARY. There are no 9, or 5, or 3 or 4.....That isn't how logic gates work.
A 3 / INT is 8 BITS...1 BYTE.
HINT: I'm right.
And you never answered my question:
"Has anyone implemented this with actual bitwise operations instead of integer packing?"
Still waiting to see these "9-bit" "bytes"."00000034".
Again, show me. There is no such thing as a 9-bit byte, that isn't how CPUs or computation work. ITS 8 BITS 1 BYTE, that is transistor / gate design architecture.
Yes 9 bits is 2 bytes. The article confusingly says 18 bits = ~2 bytes. It is the "about 2 bytes" that is confusing. They probably mean that an extra bit won't matter too much since we are bit packing the games in a contiguous stream.
BUT
In the article they don't mean that 00000034 is a bit or a byte. It is one of the possibilities and there are 495 of them and if you index each possibly in a 2 byte integer, you can decode it back to that string and get a representation of the promotions that happened in any game.
You understand me, this is most important. And thank you for explaining this exercise - but to be honest, if the article says "How to store a chess position in 26 bytes"
And you actually cannot store this in 26 bytes based on your implementation, and then you show integer bits and bytes that aren't even binary...eh.
And to be honest, like how about we store the chess position in 1 bit.
I will execute some chess position program in 1 bit, ON / 1. How about that for ultra compression? Lets just pretend those other random bits and bytes don't exist I mean (...but they do...) - it's stored somewhere else, but "HOW TO STORE A CHESS POSITION IN 1 BIT" - but ok, fine I will play "pretend" |How to store a chess position in 26 bytes (2022)|
If the rook has not ever moved yet, it gets the king's positional value. As both pieces can't overlap, assume the king's positional value is correct and the rook is at starting position.
Then, as soon as the rook is moved, it gets its actual positional value. If it moves back later, the positional value will be that of the rook's starting position (guaranteed different from the king's current positional value as the two pieces can't overlap).
It would be if castle is available, not simply if the rook has never moved.
Likewise, the position of a pawn can be assigned the king’s position if it has made the double move. You know it’s actually in the legal file and in which rank it sits after the move.
i think i just misunderstood the writing, it does explicitly say 4bits for castling. the prose around is just describing what castling is - i thought it was implying that you could determine whether castling is possible from the position of the pieces.
He starts out by using 4 bits for castling rights.
Then he introduces the other method (signify that castling is allowed by saying the rook on that side is on the same square as the king) and with that method he doesn't need any extra bits for castling rights.
Edit: it would be better on average to keep the castling bits, and omit the positions of kings and rooks if castling is possible. But that's variable length and it's simply 4 extra bits in the worst case.
$2000 will get you 30~50 tokens/s on perfectly usable quantization levels (Q4-Q5), taken from any one among the top 5 best open weights MoE models. That's not half bad and will only get better!
If you are running lightweight models like deepseek 32B. But anything more and it’ll drop. Also, costs have risen a lot in the last month for RAM and AI adjacent hardware. It’s definitely not 2k for the rig needed for 50 tokens a second
Could you explain how? I can't seem to figure it out.
DeepSeek-V3.2-Exp has 37B active parameters, GLM-4.7 and Kimi K2 have 32B active parameters.
Lets say we are dealing with Q4_K_S quantization for roughly half the size, we still need to move 16 GB 30 times per second, which requires a memory bandwidth of 480 GB/s, or maybe half that if speculative decoding works really well.
Anything GPU-based won't work for that speed, because PCIe 5 provides only 64 GB/s and $2000 can not afford enough VRAM (~256GB) for a full model.
That leaves CPU-based systems with high memory bandwidth. DDR5 would work (somewhere around 300 GB/s with 8x 4800MHz modules), but that would cost about twice as much for just the RAM alone, disregarding the rest of the system.
Can you get enough memory bandwidth out of DDR4 somehow?
Look up AMD's Strix Halo mini-PC such as GMKtec's EVO-X2. I got the one with 128GB of unified RAM (~100GB VRAM) last year for 1900€ excl. VAT; it runs like a beast especially for SOTA/near-SOTA MoE models.
reply