> This is one reason why Deepseek & GLM are priced so cheaply, they are competing with impossibly low token prices in China. They have to keep prices low, in order for people to use them.
This one does not make sense to me at all.
Deepseek and GLM are openweights, even US inference provider are selling them at much cheaper price. The price is cheap because the model is more efficient.
DeepSeek permanently cut its V4-pro API prices by 75% because they were too expensive. Without the price cut, Deepseek V4-pro tokens would have cost more than resold Opus 4.8 tokens.
Opus 4.8 is a more capable model, so almost nobody was going to pay for V4-pro at the original price.
> Without the price cut, Deepseek V4-pro tokens would have cost more than resold Opus 4.8 tokens.
You mean it's functionally as if American tokens are being price dumped in China and Chinese model providers are being forced to compete with that and innovate? So many delicious layers of irony, lol :-P
Urm, no? I man they did cut prices by 75% that part is true - but they reduced a starting price that was below sonnet.
Also it's a open weight model, doing that is impossible long term because the real price will be set by the other model providers, who priced it around 60% of sonnet inference cost. Had to look that up though, so that's today's pricing.
Is there a contradiction here? If resold Opus tokens are sold at a 93% discount, you can be a lot cheaper than Sonnet while also a lot more expensive than resold Opus tokens.
I see, After rereading the comment I was responding to I realized I probably misread/misinterpreted what they wanted to convey.
I think there isn't a contradiction and I was just confused. The price may have been discounted only to get below the price point of opus resellers. I do not have enough information on that to make any clear determination on that topic.
It's somewhat difficult to have any sympathy for Anthropic here. They're entirely responsible for selling tokens at below cost, with the age-old bait-and-switch tactic.
If they weren't doing so, then these Chinese resellers wouldn't be viable.
Radical idea, but how about they actually charge a viable price, even on subscription plans?
If resold Anthropic tokens undercut even the at-cost open-weight model tokens, because they're reselling subsidized subscription tokens, then you'd have to start selling open-weight model tokens at a loss in order to match them.
I imagine FPGA could just be part of general CPU that provides user space APIs to program them to accelerate certain work flow, in other words, this sounds like exactly JIT to me. People may program FPGA as they need to, e.g. AV1 encoder/decoder, accelerate some NN layers, or even a JS runtime, am I thinking something too wild for hardware capability or is it just the ecosystem isn't there yet to allow such flexible use cases?
Digital logic design isn't software programming, and today's FPGAs are for most intents and purposes 'single-configuration-at-a-time' devices - you can't realistically time-slice them.
The placement and routing flow of these devices is an NP-Complete problem and is relatively non-deterministic* (the exact same HDL will typically produce identical results, but even slightly different HDL can produce radically different results.)
All of these use cases you've mentioned (AV1 decoders, NN layers, but especially a JS runtime) require phenomenal amounts of physical die area, even on modern processes. CPUs will run circles around the practical die area you can afford to spare - at massively higher clock speeds - for all but the most niche of problems.
My rule of thumb is a 40x silicon area ratio between FPGA and ASIC, a clock speed that is around 5x lower. And a lot more power consumption.
If you have an application that can be done on a CPU, with lots of sequences dependencies (such as video compression/decompression), an FPGA doesn’t stand a chance compared to adding dedicated silicon area.
That’s even more so if you’d embed an FPGA on a CPU die. Intel tried it and you got a power hungry jack of all trades, master of none that nobody knew what to do with.
Xilinx MPSOC and RFSOC are successful, but their CPUs are comparatively lower performance and used as application specific orchestrators and never as a generic CPU that run traditional desktop or server software.
I have yet to see the "FPGA is less power efficient" thing to be true. People are always comparing the same circuit in an FPGA Vs an ASIC but this is a nonsensical comparison because of three reasons:
FPGAs have hard wired blocks like DSPs which do not have any power disadvantages vs "ASIC" (only advantages actually)
the likelihood that there is an ASIC that happens to implement your particular design is very low
and off the shelf ASICs like GPUs and CPUs have significant amounts of overhead for each operation. This is especially evident with CPUs. They perform a small number of operations per cycle, but they have to pay the entire fixed energy cost of caches, registers, instruction decoding, etc per cycle. This is way way worse than programmable logic if you're mostly using the DSP and the block RAM slices.
sure there are hype around MCP, but how does that bridges to "MCP Is Mostly Bullshit"?
the title is more or less click bait, even author admits it
> I am not even saying MCP is bad tech or useless. It’s just one way among others to provide context to AI assistants/agents. If you have ever built an LLM-based application, you have more or less done something similar.
but beyond just similar, it's very good to see there is a standard protocol that everyone adopts it, and the MCP will make the existing tools & services immediate ready with better models being released in the future much easier
It is really fun and sarcastic to watch all this happening. U.S. Gov. tried to block China from accessing GPU resources very hardly to stop their AI development, but actually helped China to take a leap on developing efficient and more cost-effective LLM model with constraint GPU access.
And then "China" (which is actually a bunch of super generous folks at DeepSeek) decides to release it all back to the US under a permissive MIT license.
They could've just exposed an API and kept the model to themselves but they didn't!
They could've not published their research paper, but they did, again and again - and each time they publish they discuss not just the techniques that DO work, but those that don't - saving researchers everywhere from loads of dead ends.
That is pure awesome. Thank you DeepSeek engineers for your gift to humanity.
Do they have models that try to downplay what happened on Tiananmen Square? That would be a sneaky way to shape our future in some way (and no whataboutism, we do it too).
No human is in danger of forgetting Tiananmen Square unless they didn’t know about it in the first place. Details are strewn across the Internet and in book-libraries all over the world. New generations of students and interested kids can easily learn about them.
Additionally it has been shown that making models forget things lobotomizes them, so no SOTA model can ever do that and be SOTA. They might be post-trained into pretending not to know, but the technology fundamentally cannot resist jailbreaking.
Do you have examples of knowledge that has actually become at risk as a result of this one AI model being added to the pile??
I doubt you ur GPU sanctions have had much of an influence one way or the other. They can get their resources from third countries even if they can’t get them directly from the USA. I wonder if the USA will eventually try to lock down higher end NVIDIA GPUs and prevent export all together.
from Company perspective, I think the answer is yes
but from individual perspective I don't think that is the case. Since AlphaGo first time was released and beat world-class players, have all these players gone? not really, but it even promotes more people study Go with AI instead.
As a software engineer myself, am I enjoy using AI for coding? yes, for many trivial and repetitive works, but do I want it to take my coding works fully away such that I can just type natural languages? My answer is no. I still need the dopamine hit from coding myself, either for work or my for hobby time, even I am rebuilding some wheels other folks have already built, and I believe many of ours are the same.
The guy that got beaten literally decided to retire immediately after explicitly because AI displaced him
> On 19 November 2019, Lee announced his retirement from professional play, stating that he could never be the top overall player of Go due to the increasing dominance of AI.
I do get your point though that overall player count is still fine
I can imagine how depress Lee felt when being beaten by AI the first time, but looking at the bright side, we see Shin Jin-seo as the rising star in the AI-era of Go by leveraing AI to help him training.
if U.S. could have opened its market to Chinese EVs (yes yes I know Chinese Gov subsidizes those vendors a lot, but U.S. can still do punitive tariffs, instead of just ban[1], and of course, making excuses out of national security, meanwhile Tesla is doing okay in China) these U.S. vendors have no competitions is the reason why they are far behind.
There are a few Chinese EVs available widely in Australia (BYD, MG, and recently Zeekr) that all conform to Australian Design Rules (ADR), which have been "harmonised" with the European ones for a while. So, most likely this would get the vehicles pretty close to passing the US safety rules, though they'd still have to go through testing and certification.
This one does not make sense to me at all.
Deepseek and GLM are openweights, even US inference provider are selling them at much cheaper price. The price is cheap because the model is more efficient.