This is what the far right in Brazil wants you to believe, but it's not true. It's especially not true in regards to X, as they have repeatedly ignored judicial decisions that are very similar to those by European courts (ie, remove illegal content).
Well, one of the people calling out this insanity is Glenn Greenwald. How can he be "the far right in Brazil", when his reporting on Lava Jato is one of the main reasons Lula is not in jail any longer?
I admire Glenn, he did a great job reporting on Snowden revelations and the Vaza-Jato scandal in Brazil. I believe his stand is principled, but he appears to have the shortcoming of often trying to apply concepts from the US legal and institutional framework to places where it simply doesn't apply, like trying to fit square pegs to round holes. This sent him into a wild goose chase in this case.
Greenwald is an odd one, I'll give you that. He seems to see this issue from a libertarian perspective. He would also say that Germany banning Nazi speech is censorship, for example. It's worth nothing that his last professional occupation was being a stooge in a far right show.
How do you know that they’re similar? As the source article describes, the court has not published the orders they’re asking Twitter to enforce, and they don’t seem to be available at all to the public.
...how is that different to EU court rulings? Take your comment, and swap out "Brazil" with the name of any other country, and absolutely nothing changes.
This is partially the reason why we see LLM's "plateauing" in the benchmarks. For the lmsys Arena, for example, LLM's are simply judged on whether the user liked the answer or not. Truth is a secondary part of that process, as are many other things that perhaps humans are not very good at evaluating. There is a limit to the capacity and value of having LLM's chase RLHF as a reward function. As Karpathy says here, we could even argue that it is counter productive to build a system based on human opinion, especially if we want the system to surpass us.
RLHF really isn't the problem as far as surpassing human capability - language models trained to mimic human responses are fundamentally not going to do anything other than mimic human responses, regardless of how you fine-tune them for the specific type of human responses you do or don't like.
If you want to exceed human intelligence, then design architectures for intelligence, not for copying humans!
Intel seems 100% tied to its outcomes as a foundry now. If there is any disruption in chip prodution from Taiwan for example, I can see its stock doubling in price.
Devils Advocate: If there is no disruption in Taiwan, and TSMC continues to execute as it has, and Intel continues to execute as is has (not), then Intel might go to $0 or survive only because of government subsidies and military contracts.
NVIDIA, AMD, Apple and other companies seeking cutting edge performance will choose the most advanced node to stay competitive. Being the second best foundry has historically not been good business.
TSMC today has more than 60% of the foundry market share, and an estimated 80%+ market share for the leading edge.
If you exclude Intel themselves (although they also use TSMC now) and Samsung, TSMC is pretty much the sole supplier to the leading edge.
So, yes, TSMC has the capacity.
Your Boeing / Airbus analogy hinges on multiple factors. First, Airbus and Boeing have comparable capacity. Second, they have comparable products. Both are not true when you compare Intel and TSMC.
NVIDIA is heavily supply constrained. Why haven’t they sourced Intel or Samsung as second source?
Historically, there are also many other issues with Intel operating as foundry. Do you think NVIDIA and AMD will be happy to send their CPU and GPU designs to Intel for manufacturing? Independence was one of the main drivers why TSMC was founded. To have an independent supplier who does not compete with its customers.
It was a genuine question, I don’t follow the sector closely.
That said to poke at your answer a bit:
If nvidia is heavily supply constrained, wildly profitable, and strictly limited to TSMC doesn’t that dynamic lead to other customers being pushed aside? If so, doesn’t that open an opportunity for Intel if they can execute?
The immediate winners will be Samsung and Globalfoundries, since they have wide arrays of foundries with different processes (including specialized ones for automotive/avionics grade components, dram, cmos sensors, etc.), spare capacities, existing customers with designs tailored to their processes, and extensive experience in helping customers getting their designs ported.
Intel is going to fight an uphill battle with all these issues, it's likely that TSMC the other competitors can expand capacity faster than Intel can convince customers to gamble on them.
I would not. There is nothing to indicate Intel leadership is capable of executing. Bean counters are not the kind of leadership you need in a changing environment.
Isn't the whole problem that it wasn't QA'd out precisely because Intel no longer know how to ship good products? Sure Intel have the resources to make their customers whole on this screw up but the concern is that Intel have had a pattern of poor execution in the fast 10-15 years and this is just the most recent example.
One of the videos from the well connected tech youtubers (I think it was GamersNexus) said that QA had caught the issue (if you push this bus too hard with voltage, there will be issues) but management ignored it because they needed to compete with AMD. There was also something about pushing previously binned i7 parts out as i9's by pushing the voltages to meet demand too. I'll see if I can find the reference a bit later.
There are few absolutes so stop trying to reduce the problem to one such as "Intel bad" because that's disingenuous.
It's really hard to reason about things until you hit large quantities of hardware in production, separate statistical failures from non-statistical failures, separate end user idiocy from real issues and do the analysis, then reproduce it conclusively. This is not helped by the target end of the market demand being interested in running things close to the line on boards kicked out by the lowest bidder. On top of that you have large vendors like Dell and Lenovo who take forever to feed information back on warranty claims because their end user support is fucking awful.
Basically it's very hard to reason about this pre-production despite every effort and QA step in the book being followed.
I am not reducing it to "Intel bad". There's tons of evidence of them having execution problems in recent years. Is it savable? Possibly. Intel is huge and has a good amount of time and support so they certainly have the resources to do so.
The available evidence that they are actually doing the right things to turn the ship around isn't that positive in my, or apparently many very large shareholders view though.
Intel publishes much more information about their future roadmap than any other company in this industry.
During the last couple of years they have achieved in time all the product and process launches that had been announced with years in advance, up to Meteor Lake on the new Intel 4 CMOS process at the end of 2023 and Sierra Forest on the new Intel 3 CMOS process at the middle of 2024.
The only things that could be considered as under-achievements are the facts that in the Intel 4 process the maximum clock frequencies have been lower than hoped and the production yields are still low. Nevertheless, these 2 problems are not surprises. Every new CMOS process launched by Intel during the last decade has been affected exactly by these 2 problems, at least for its first year. For each such process, Intel has succeeded eventually to solve the problems, but they have always needed the experience of multiple years of production for that. With Intel 4 this will not happen, because it will not be used again for another product, being replaced by Intel 3, which probably incorporates whatever fixes were suggested by the experience with Intel 4.
So the financial problems of Intel are not caused by Intel failing to execute their roadmap.
Nevertheless, the bad financial results have been perfectly predictable from the Intel roadmap, as it had been known for years, because that roadmap does not attempt to achieve parity with the AMD server CPUs before the end of 2026 at the earliest.
The financial results of Intel have shown profits for the client CPUs. They have shown losses for their foundry, which have been unavoidable, due to their need to invest huge amounts of money to recover Intel's handicap vs. TSMC and due to the need to transfer a part of their chip production to TSMC, while they are developing their new processes.
They have also shown great losses for the server CPUs. The losses in server CPUs are likely to be caused by Intel having to cut their prices a lot when selling to big customers, in order to convince them to not buy the superior AMD server CPUs.
It is very likely that Intel will continue to have great losses in the foundry and in the server CPUs for at least one more year, based on their public roadmap. After that, their results will depend on how good the Intel 18A CMOS process will be in comparison with the competing CMOS processes of TSMC.
The Intel roadmap for manufacturing processes and for client CPUs has been decent. While some of the Intel choices may be undesirable for their customers, other choices would not have improved their financial results.
Where Intel could have conceived a much better roadmap is in server CPUs, where they have their greatest losses. When AMD launches a new microarchitecture, like now with Zen 5, after 3 or 4 months they also launch their server CPUs with that microarchitecture, like with AMD Turin later this year. On the other hand, when Intel launches a new microarchitecture in their client CPUs, like in the next couple of months with Lunar Lake and with Arrow Lake S, they follow with server CPUs using the same up-to-date microarchitecture usually only one or two years later, e.g. in the second half of 2026 in this case.
While the Lion Cove and Skymont cores of Lunar Lake and Arrow Lake S are competitive with Zen 5, the Intel server CPUs launched these days, Sierra Forest and Granite Rapids, continue to use cores that are only marginally better than the cores used in Alder Lake three years ago.
The new Granite Rapids server CPUs will be able to match the number of cores per socket of the AMD server CPUs, for the first time after many years. However, they will use cores not much better than Zen 4, which will be inferior to Zen 5.
This situation would have been easily avoided by Intel if they would have launched now only CPUs using their up-to-date cores, instead of using obsolete cores in their server CPUs.
The reason why AMD does not have this problem is because they design a single CPU chiplet that will be used both in their desktop and high-power laptop CPUs and in their server CPUs. On the other hand, Intel duplicates their design and validation efforts by designing distinct chips for desktops and for servers. When AMD finishes the design of a chiplet, it is ready for both desktops and servers. Their server CPUs use a few months of additional testing, but that is all of the delay. It is not clear how the design work is organized at Intel, but it appears that there is not enough overlap in the design of the client CPUs and of the server CPUs, so the latter are finished only after one extra year or two.
Could you please stop posting unsubstantive and/or flamebait comments? and also turn down the general grumpiness? You've unfortunately been doing these things repeatedly. They aren't what this site is for and destroy what it is for.
It's very widespread. But if you look past the stock trade and the pop-psychology, it's pretty clear that Intel today is on a different trajectory than they were 10 years ago. In that time they've changed CEOs, reframed their roadmap, kickstarted auxiliary businesses and invested in RISC development. The catastrophic potential of Intel losing desktop market share was overstated and over-anticipated. Now, cutting off their desktop segment is practically the most profitable route.
HN has a habit of jumping the gun on financials without accounting for what the market is doing or where things are going. Intel doesn't have the dominant position they used to, but their trajectory is heads-I-win-tails-you-lose levels of redundant. Their roadmap is the good kind of boring, and they're the only fab I know that has TSMC's flagship products in the crosshairs. We're desperately trying to declare Intel dead while already anticipating their resurrection; that suggests to me there's a good pulse left.
I expect the problem is more that people disregard fundamental truths because they think macroeconomics supersedes realpolitik. It's a particularly HN-oriented tradition that leaves a lot of people with sore-looking opinions after the fact.
But boring stuff has low margins. Intel has been living in a high-margin world at least since the 8086. They currently don't know how to survive on low margins. I'm not sure they can learn.
This is what a lot of people who have no idea about the semi industry miss because they get all their education from news headlines. Which is what I assume leads to Intel's massive stock drop. People voting on FUD, kind of like the opposite of the hype that pushed Nvidia to the moon, but when you asked investors who bought Nvidia, what Nvidia actually makes, most are clueless and only bought out of hype and FOMO.
Everyone assumes the world only needs 3nm chips because that's what TSMCs latest cutting edge node is and that's what they learned powers their iPhone, but the reality is the world runs on nodes much larger than that. To put this into perspective, the new TSMC fab being built in Germany will tape out 28/22nm nodes and then upgrade to 16/12 nm nodes mostly for robotics, industrial, IoT and especially automotive. Just let that sink in. Even Global Foundries is stuck at 12nm.
Intel's 4 process is cutting edge by comparison and plenty of international players who can't afford TSMC or Samsung would gladly pick up that node capacity for the right price.
Yh, 28+16 is basically pretty much all you need for industrial semis. The other key part I think is basically going to be military, I've watched an interview with one of the big wigs at Baikal and he claimed that the Russian miltech industry would be fully self-sustaining if there was domestic 28nm capacity.
I remember one comment from during covid when the supply chain was messed up (possibly from Intel) saying productivity could be improved a lot for non-leading edge chips by getting designs off relative ancient processes to relatively recent, which would let the industry modernize old fabrication plants. It requires work for that new design though, and moving away from thoroughly tried and tested
That was quite naive. The problem is there is little motivation to do that. The designs can't just be scaled down, they need to be recompiled from scratch with the target process' standard cells and verification. That work is expensive as fuck. Then there's packaging and managing the supply chain and data around it. Why bother when demand causes prices to go up? Just pocket the money and wait. If you spend that on more capacity and reengineering you just burn the money.
I never said what you're claiming. Please read comments carefully, assume good faith in the comments you read, and don't put words in my mouth.
Like I said, Intel's layoffs are a reaction to their stock tanking, and their stock tanking is due to investor behavior which is sometimes rational, but a lot of the times not.
The graphs seem to indicate their model trades blows with Llama 3.1 405B, which has more than 3x the number of tokens and (presumably) a much bigger compute budget. It's kind of baffling if this is confirmed.
Apparently Llama 3.1 relied on artificial data, would be very curious about the type of data that Mistral uses.
This article is borderline comical. The thin list of his accomplishments focus on increasing revenue ($15B to $70B) and launching or acquiring successful initiatives (Xbox, Skype, Azure). Then the rest is a recollection of his biggest embarassments, like terrible mobile products, misguided Windows strategy and trailing AWS by 7 years.
Reporting on business issues is always muddled by a lack of proper comparisons, along with cherry picking. For example, this article makes the argument that increasing Microsoft's revenue by 4x was very impressive, even though the stock value stagnated. However, when evaluting his tenure as owner of a basketball club, he is declared successful because its value doubled. The problem is that Microsoft was eclipsed compared to its peers at the time - Google, Amazon, etc. -, and likewise the average basketball club doubled in value as well.
It sounds like Andrej's ambitions stretch beyond AI/LLM courses too - but they're a natural "first course" starting point because that's where his own teaching expertise is focused.
It is. Though they are also doing some really smart and thoughtful stuff with LLMs to power courses and learning environments (https://www.khanmigo.ai/ is a part of it).
The author is making the point that Alphafold 3 is not so impressive - it is simply regurgitating its train set, and it's not so good for inference.
I think his central point is fair and interesting. The test train split is apparently legit, as they used structures released before 2021 for training and the rest for testing. However, there was no real check for duplicates, and the success rate might be inflated by a bunch of "me too", low hanging fruit structures that are very slight variations from what we know.
However, I'm not sure I agree with his skepticism. LLMs suffer from the exact same problems - getting it to write a Snake game in any language is trivial, but it is almost certainly regurgitating - , but can be useful as well. I mean, if for various reasons people are publishing very similar structures out there, there's certainly value in speeding up or reducing that work considerably.
It would be simply stupid that we got so much worse at this technique in the past 10 years. You are welcome to post a link to your data source here, you already had two opportunities to do so.
- Vast cost reduction (>10x)
- Performance parity of several open source models to GPT4, including some with far fewer parameters
- Much better performance, much larger context window in state-of-the-art closed source LLMs (Claude 3.5 Sonnet)
- Multimodality (audio and vision)
- Prototypes for semi-autonomous agents and chain-of-thought architectures showing promising avenues for progress