Since no one specifically answered your question yet, yes, you should be able to get usable performance. A Q4_K_M GGUF of DeepSeek-R1 is 404GB. This is a 671B MoE that "only" has 37B activations per pass. You'd probably expect in the ballpark of 20-30 tok/s (depends on how much actually MBW can be utilized) for text generation.
From my napkin math, the M3 Ultra TFLOPs is still relatively low (around 43 FP16 TFLOPs?), but it should be more than enough to handle bs=1 token generation (should be way <10 FLOPs/byte for inference). Now as far is its prefill/prompt processing speed... well, that's another matter.
Instead of each node storing the next- and previous-pointers separately, store a single pointer which is the XOR of the two. Which is obviously an invalid pointer. But when iterating, XOR the previous node's pointer with the combined pointer to get the next node's pointer, and so on. You can iterate this way in both directions. Feels illegal. :)
Does it? How many people skipped getting rich because they could have been richer? Any factual examples?
BTW, rich don't actually pay much taxes. The luxury life they live is usually not taxed, most of the things they do is considered business expense.
When a worker flies to Ibiza they first pay social security and income taxes, then they pay consumption taxes like VAT.
When a businessman flies to Ibiza they deduct whatever they can as an expense so they don't pay income tax and VAT. For whatever they can't claim that it is a business expense they will pay with a cheap loan against their assets and avoid paying income taxes. Since they still have those assents, they pay just the interest later when the assents increase in value. If their business fails those assets fail, the bank takes the assets and no taxation happens.
I shorted Royal Caribbean on the way down and am now buying shares. I figure most of the impact is priced in, I am a long hold on them now and we will see where it shakes out when it is all over. They have an extremely loyal customer base.
Thing is, daos have potential intrinsic value in the way they decentralize governance.
An NFT would only be valuable if it constituted an enforceable contract over some share of future Revenue from use of the object of the contract. Right now, that's not true, thus NFTs are not valuable, cuz they don't earn anything (unless you find a bigger sucker -- but that's speculation, not investing)
Lots of people want to hop on the crypto bandwagon and use it for everything...it's better to use crypto for the things it is really good at: immutability, availability (maybe), nonrepudiation definitely. Confidentiality? Not really...
> if you invested at the wrong time in the SP500 20 years is not enough to recover.
That's simply not correct.
And important note here is that I'm obviously not advocating for investing 1 million dollars in a single SP500 ETF. For some people that may very well be the right thing, but if all you've got is 1 million dollars, you're taking a ridiculous risk putting it all into one thing, because what happens if some years down the line, stuff happens in your life, and you need to take out a good chunk of cash and it coincides with a terrible market, not a stellar combo.
So obviously basic diversification is critical, based on your personal situation. Can be accomplished easily with 3-5 cheap ETF's. Buy and hold for many years, until your life changes.
But as for SP500.. You could have invested in virtually any SP500 high, and as long as you can hang on for more than 10 years, you can almost not lose. Quite the opposite, sitting in cash is sure death, because you're gonna miss all the highs, sure the market may drop 30%, but that dip could have been in a period of 40% gain.. And you're losing out on the most essential part of making money long term: compounding interest. So.. Just invest, there's (almost) no bad time.
As for OP and timing.. I wouldn't know.. Way too many unknowns to give any kind of specific advice.
Someone else linked to the "three fund portfolio" and that's close to what most people should do. Get VTI, VXUS and BND in some ratio based on your age, personal situation, investment horizon etc, and then just let it sit.
You have it backwards. The MIT/Expat does required you to retain a copyright notice (in the form of a copy of the license, not in the form of file headers), which Zen does. In addition, creating a derivative work does in fact give you copyright on the new work. The Shakespeare comparison does not apply because Shakespeare is not licensed under the MIT/Expat license.
With the important caveats of 1) the implicit costs of murder-for-hire are probably pretty high (risk of going to jail, e.g.) and 2) killing someone doesn't necessarily make the lawsuit go away—though that doesn't help the person being killed.
So the utility formula is something like:
(p[c] * c[f] + c[a]) ÷ p[l] > c[l]
where:
p[c] = probability of getting caught
p[l] = probability of the lawsuit being dropped
c[f] = value of your freedom
c[a] = cost of an assassin
c[l] = cost of losing the lawsuit
You can layer in additional probabilities—like the probability of the assassination being successful, or the probability of winning the case. This is just fun to talk about the utility formula of assassinations on HN.
But it's probably true that as c[l] increases, c[f] increases as well, and p[l] drops. So it's rarely (never?) worth this trade-off.
One of my minor goals for 2020 is to switch all the gear I carry around to being USB-C PD. I've been doing some research on the subject, so just thought I'd throw my 2 cents in (some of this stuff mentioned in other comments).
I started off the year with an Innergie 60C, which has worked great for everything I've thrown at it and is tiny (55mL volume, 88g, 60W USB-C PD). It was on-sale recently, but even at full price has been worth it for me. (Although the newer RavPower PD Pioneer 61W is a good alternative that's almost as small and 1/3 the price.)
Instead of the Anker PD1 18W, for a small charger, I use the Anker PIQ 30W - it's very slim, which actually makes it much easier to carry around. Also, the IMO folding plugs (and extra juice) are definitely worth the extra volume.
Recently, for my Mavic Pro 2, I just got a cheap Cablcc USB-C to Plug Receptacle Charger cable and a RCGEEK car charger that seems to work well as a nice 2-in-1 (it only saves a tiny bit of volume and weight over the wall charging but it also provides car charging).
Also, although I'm unsure of how often I'll need it, I got a 100Wh Zendure Super Tank power bank (which provides USB-C PD 100W charging), since it was discounted heavily for Black Friday (it was only $100 on sale vs $200+ for other similar options).
Also, I travel internationally a lot, and the MOGICS Donut has been my go-to combination power strip and adapter - it's the best/most compact device I've found (and has 2 USB-A chargers built in to handly legacy devices). https://www.mogics.com/3824-2
are you expecting a higher stock price? because their 90 PE TTM doesn't look like it's in line with the rest of the industry.
don't get me wrong, i was also long AMD, and still partly am, but i don't think the stock can go higher than 36-37, or even maintain that level long term, without some some mega quarters. which according to their own estimates will not be the case.
I don't think GraphQL is over-hyped at all. Maybe it's flawed, but the design is absolutely on the right traack. GraphQL completely changes how you work with APIs in a front end.
I work on React apps, and by using GraphQL, a component's data requirements can now be entirely declarative. For example, a component can do this (simplified):
The component knows what it needs to render, so it declares that. With TypeScript, you can get type-safety all the way from the backend to the frontend, which means your IDE (e.g. VS Code) can correctly autocomplete, say, "creator." and suggest "name". It's rather magical.
I work on a product called Sanity [1], which evolves GraphQL one step further. It's a Firebase-like data store with schemas and joins, and a structured data model. Without implementing anything on the server end, you can run queries like these:
This doesn't just follow a child object (creator), it also joins with an unrelated set of objects (topPosts, which finds the top 10 most viewed posts created by the same user) as well as joining a bunch of 1:1 relations. I'm totally biased, but I think it's a game-changer when it comes to writing web apps, because you can just dump the whole query in a React component, no state management or API writing required.
The idea that democracy and freedom are somehow foreign to Chinese people and can't work in their culture is a myth peddled by the CCP to serve their own interests. Democracy and freedom are flourishing in Taiwan. They are flourishing in South Korea (who aren't Chinese but whose traditions are no more democratic than China's). They're doing OK in Japan.
"The Chinese point of view" is an ambiguous term that the CCP uses to its advantage. In fact it is the Chinese Communist Party's point of view, which they have managed to indoctrinate into most Chinese citizens via their control of all Chinese media. Where Chinese people have escaped CCP influence, they tend to have quite a different point of view.
Let’s replace China with X so there is no xenophobic rebuttals.
X is an authoritarian regime that has no rule of law, president has self declared perpetual status in the office, piracy is rampant, no respect for privacy of others, there is an app called ourchat that is effectively owned by the government and is increasing becoming a necessity, no media let alone any kind of investigative journalism especially against the government, your social score goes down if you buy a particular book, you cannot sue the government or even think about it, punishment can include selling your organs for arbitrary reasons, the list goes on and on.
If X were an impoverished country like Somalia, the tune would change and most people would condemn such a society. I want to do so fearlessly but sometimes people see it as an attack against the Chinese people. I’ve been to China and have spent many months there, made lifelong relations, etc. I have no room for any concession or bargain for the argument that authoritarian rule has benefits - yes it does but at aforementioned costs. China has risen above due to government’s iron grip over every aspect of the country. It is doing so at a cost. Fundamentals don’t change even if one sees the strategy panning out. An eagle in the world of doves can kill a lot of doves and have short term evolutionary imbalance. But soon, the marginal cost of turning into an eagle is so small so there are new eagles popping up in the population all of a sudden. This balance oscillates in the short term, but evolutionary pressure returns it back to an equilibrium. Fundamentals of eagle and dove dynamics don’t change even though the state of this system shows “success”.
I’m in the position to criticize any authoritarian regimes in the strongest way possible - be it China or any other country, it doesn’t matter. I don’t want to die seeing this world turn into a power grab for a few with a consequence of a dystopian society. I wish the next superpower would be a country such as Norway or Sweden, it would set such a utopian example for the world to move into the right direction.
Dude, you can strap your smartphone to a quadcopter and you have a flying 20-megapixel 60 fps camera with a multi-megabit internet connection. You can do real-time control of the flight. That's something you couldn't do with a 386, not unless you had a Predator to put it in.
And there's a whole category of new challenges not because of limited resources but because of abundant resources. Your photogrammetry drone can generate 72 gigapixels per minute of photogrammetry data. You have 200 teraflops on your desktop. How good of a 3-D model can you make? With how little human effort?
How about making things simpler and more flexible? Old GUI toolkits were designed around the need for 2-D raster operations to be hardware-accelerated. Are there simpler designs possible now that that's no longer a constraint? I'm exploring this in BubbleOS.
Sure, lots of game AIs are finite state machines. So in lots of games there's no challenge unless the NPCs gang up on you. But AlphaGo is also a game AI. It's a bit more sophisticated than an A* search! What would a game look like where you worked on a team with such AIs?
Smartphones also have multitouch. Yet >90% of people's interaction on them is using an on-screen keyboard, one-finger scrolling of lists, and tapping on prepackaged options. Can we do better? Are there UI paradigms that multitouch enables that would allow more creativity, despite the horrifying levels of lag in existing systems?
Security is a big problem, and most of the world is wasting their time on approaches like virus-scanners that can't work even in theory. But then there's seL4. What would a personal computer based on seL4 look like? How could we translate the guarantees it provides into practically useful power in the hands of everyday people?
Can you do voice recognition on every FM and AM radio channel in your area at once? What's the minimum hardware you'd need to do it?
Maybe there's a third option aside from continue the state quo or convert fully into a democracy? Maybe the party can evolve into something that's more and more benevolent to the people it governs?
As a Chinese who emigrated because of the many social problems of China, I'm not sure switching to democracy right now will be beneficial for regular Chinese. You've all heard stories about how bad mannered Chinese tourists are, but those are still somehow better than the average Chinese I'd say. Think about people who still eat shark fins today, who have no problem buying rhino horns or pangolin skins. (there are more extreme stuff, one example: https://www.animalsasia.org/intl/media/news/news-archive/fiv...). If you ever visited mainland China you probably noticed many public restrooms have no toilet paper. Some places tried to put toilet paper in but people would steal all of them right after.
I am not sure democracy can work with people like that. I fear it wont. I fear not enough educated/informed people will make democratic decisions that are harmful to themselves. I fear a democratic China will become the next Russia, the next Brazil, the next Turkey. At least the current party has done more poverty fighting than most of the third world democratic countries. (data available on world bank website)
Does China have problems? Absolutely. The question is if you really understand what are the problems, before we even start talking if the proposed solutions will work or not.
>Will the PRC just continue to exist as it is, forever? My last remaining hope is for massive power grabs by xi jinping (already happening), followed by government ineptitude as he ages and underlings squabble, followed by chaos upon his death. Beyond that I can't see any way out.
If that happens, is that a good thing? To you, or some Americans who see China as a foe, maybe. I doubt such a chaos is what the average Chinese wants.
>What you have instead is knowledge debt as engineers turn over ... This isn’t tech debt
You know, if there only were a way to pass knowledge from one person to another aside from directly talking to that person. Perhaps if we had means of preserving knowledge, and even making it a part of the job description... But alas.
Joking aside, not writing technical documentation and product design docs is very much technical debt.
A lot of information that is lost with turnover is information that could have been written down and passed on.
Documentation is not jut code comments. It is:
-A solid bug tracking system, with clean and reproducible bug descriptions
-A version control history, where every change has a why in addition to a brief how, and which links with the bug tracker
-Well-written tests, which not only verify how a system works, but showcase how to use the APIs.
-Documentation for tests explaining why this input was chosen, what kind of output is expected, and why it was written that way
-High level comments in code (especially libraries) explaining how the API should be used
-High level documentation explaining the structure of the code that is checked in, and lives with the code. The documentation can be in Markdown or a plaintext README, but should be there
-A "code map" telling you what the code in various directories does, and where to look for things.
-And, most importantly, a lot of design docs - for every feature and algorithm. "We are solving this problem. Here are the alternatives considered. Here is what we are going to do and why. Here is how we'll implement this. We used ideas from this, this, and this paper".
Yes, all of that is needed. The price you pay is engineering time. But if you take on the debt, your new engineers will have Hours Of Fun[1] trying to figure out, say, whether the imaginary nanometers[2] model parameter increasing as a result of their change is a good or a bad thing.
What I am saying is documentation debt is technical debt. Documentation is just as important as the code. And having it does help[3].
You might think that your code doesn't need all of that. And that's fine; but that means you plan on never paying that debt. Maybe you plan for your startup to fail/sell out in a few years. Maybe you plan to leave that job. But anything long term means you have to write things down.
And if you haven't been doing so - now is as good time as ever to start.
[2]True story. There was a good reason for why the distance parameter had complex-valued nanometers as units; it was an elegant way to do improve a model while staying backwards-compatible. It was written down nowhere, and have fun googling "complex nanometers". They only exist in that code, AFAIK.
https://purebulk.com/search?type=product&q=magnesium
Or grab these next time in Europe. 0.50 Euros effervescent tablets https://www.rossmann.de/de/gesundheit-altapharma-brausetable...
I never heard about the B12 problem. But effervescent tablets without artificial sweetener are unheard off.