HN user
Yeah that's all hackery I think, if you read at the specs there is always a device involved, bitwarden and company just pretend to be a device or have an extension that just ignores the spec.
Cut and Paste is three operations.
Cut is two operations copy and delete, copy is never undone by undo it should not be undone when you cut either. I cut, undo, ..., undo, and paste multiple times a day. It's a feature not a bug.
Cut & Paste is not atomic
Yes because it's two different actions.
What does "Ghost Cut" do if you paste multiple times? Paste the cut text first and then what? The previous thing in your clipboard? Why does my editor need to read my clipboard if I am not pasting (to implement the rollback)? What if there was a secret key in there do we just hand it to copilot or whatever extension is running?
Cut is copy and delete plain and simple.
Cut and paste is a poor analogy in the file explorer, what the file explorer does is a move, "cutting" fills the first parameter and "pasting" fills the second. Hence the greying out and not doing anything until you paste. There is no clipboard for the filesystem. Also in the file browser it's extremely unlikely you would want to paste into multiple places, something not true in text editors.
EDIT:
Now listen I am not saying the proposed semantics are bad, to each their own, but it's a different operation all together the clipboard is not even required you could have a separate short cut that grays out the text and then moves it to where you want. The whole thing would be atomic called a move and be cleaner in all ways.
My issue is not the lack of Linux support, my issue is the dependence on hardware components, that are distributed by a small group of incumbents, that is pretty much baked into the spec.
What they want is to lock your identity to your android and/or iphone devices.
Unpopular opinion but correct the whole thing has been designed to lock you to devices they make and have themselves be the arbiter of your authentication.
If that wasn't the intent they could have make the thing work like ssh keys, encrypted at rest, you can take them wherever you want.
I think the intended workflow is you login with your phone and that device is now the authority that allows other devices to issue their own passkeys.
In my opinion it's a bad plan, because it elevates certain devices to privileged status, if you lose your phone you are hosed.
Passkeys should be allowed to be synced between devices and stored on password managers in the cloud. I am making my own password manager for my personal use, but have not delved into passkeys.
They probably could, but if done on a large scale, they are likely to be detected.
Oh yeah the may is on the 2 year time horizon. It could be 3 or 4. Or next year.
whoever burns their models to ASICs fastest.
There is already custom hardware see cerebras.
GPUs have a lot of slack there is at least one lab that had a (small 8b) model generate almost 3000 tokens per second on a MI300X for a talk, instead of the typical software stack that did maybe 100ish tokens per second.
High bandwidth flash storage is in the works, i.e hard drives with TBs of storage and over 1 TB per second of read speeds. Meaning that in a couple of years you may be able to buy a card with 40-90GBs of HBM and 4TB of HBF and run a 3T model locally at a reasonable speed for 10-20k as opposed to a cool mil.
Of course, as a rational person, you wanted to side with the scientists; if nothing else, it seemed like the winning move.
The rational person does not engage with the argument. Because there is nothing to be gained from the argument.
The winning move is not to play.
A typical server that costs 10k to 30k to own and operate can serve between hundreds and thousands of requests per second of a traditional web application like facebook for 2-4 kW of power, the marginal cost of each request is effectively zero.
A single response from kimi k3 requires hardware that cost between 500k and 1m dollars up front and draw over 20kW. Each request costs at least 5% to 10% of the charged cost.
Am I the only one who cant read the hidden message?
And AI is not new, in this regard. The flattening of all pains into a total loss of pain has previously been the job of recreational drug use or theology
Son, my dishwasher is not a religion, although I may declare holy war upon you if you try to take it from me.
I run 3.6 but yeah you are right
I run qwen 27b at home when working it pulls around 400W. I get 40ish tokens per second generation and more importantly about 1000 tokens per second prompt processing.
In an hour it can process 3.6 million tokens or generate 144000 tokens. This costs me about 15 cents given my electricity prices.
For sonnet the equivalent token costs are 7.2 dollars for the prompt processing or 1.4 dollars for the generation. The cloud is 10x more expensive for generation and close to 50 times more expensive for processing.
The 5.2 tokens per second generation is not that bad, what kills it is the 16.2 prompt processing that makes this too slow to consider even if you have the hardware lying around.
Broken clocks and all.
My 2c, stop looking for excuses, stop assigning your self worth to the quality of your work.
Mistakes happen, bugs are impossible to avoid. You may need to add rigid processes to your work to avoid the most egregious examples, but you just have to live with the fact that you will write bugs.
If you find you write more bugs than your peers and they have more impact than your peers, that's fine maybe you are not cut out to be a "systems engineer" maybe move to something more forgiving like frontend or something, you will probably be happier.
There is a reason that I am not working on avionics or respirator firmware. I don't have the discipline to follow the processes required to minimize the chance of accidentally killing people and I don't want the legal liability.
You don't have to be working on "important" stuff, John Carmack one of the best and most celebrated developer of our time spent most of his career working on games.
Be mercenary, do not take pride in your work, do your work for money. You will be happier, take pride in who you are and what you do outside of work.
I believe the scaling comes in later, to turn the 1 and -1 into large numbers that may or may not activate the next layer.
The way they do it is packing like the other comment says.
Each byte represents 5 trinary values instead of 8 binary, and there is a little bit of waste.
The article misses the point of writing.
Yeah I have found that AI in general has a procrastination nullifying effect.
Before dealing with anything that might put me off. I can just ask the agent to do it for me. And then, do something else, take that break, but regardless in a few minutes I will have something to jump on instead of the same blank terminal with the same blinking cursor judging me. It really makes taking the first step, much easier and then the ball just gets rolling.
I see what his point is to be honest though, it's easy to say just one more week of polish, just 5 more features, etc.
No matter how much you don't believe there is a tiger behind the bush. The tiger really believes you are going to be tasty.
The reason it talks that way is clearly am attempt to hook into your dopamine system.
If what you told it to do is 'load bearing' then its important.
'You are absolutely right', because you are a smart fellow.
'Honest take', because it's being honest with you because it trusts you and you should do the same.
My 'honest take' these are absolutely garbage patterns that have no place in an session interacting with AI.
1. 'Load bearing' is a figure of speech that bears no loads.
2. 'You are absolutely right' it's not the agents job to judge that, it's job is to do what I told it to do.
3. 'Honest take', so everything else was not honest? Absolute honesty should be the default and is implied.
These words add nothing to the task at hand they are a poor attempt to hook you into using this particular model.
In my view the article makes the following points.
1. AI does not need to get better for the MBA to use it instead of your labor.
2. They don't need to be good at using AI because they don't care about quality.
3. Most technical problems are unsolved right now and AI won't change that.
OVH is much more developed on the managed side, they even have managed k8s as a service.
And you know their infra generally works, you know when it's not on fire.
And when it is, well you are not going to be the only one down.
Asking claude to implement a single feature that takes under 30 minutes consumes 10-30 dollars of tokens in api costs.
Speed is really important because at these speeds they can probably intercept drones flying with low cost jet engines. So it adds a huge capability.
Oh and probably loiter time, the rocket is you shoot and it's gone. The interceptor may have enough battery to fly around for a few minutes looking for the next one if its target is downed by something else. Plus you can probably recover these if they miss. This is all very speculative though.
Maneuverability not so much modern AA missiles can pull over 60gs. That capability costs 100s of thousands of dollars though.
Cost, your average stinger cost 38000 dollars in the 80s. I am guessing here but they are aiming at a price of probably under 10000 dollars.
Now why doesn't anyone take a rocket and stick the drone guidance on it? Again I am only guessing here, the drone guidance components probably can't cope with 2-3 mach. At 1000 meters per second with a 60 fps camera you advance 16.6 meters per frame, add to that the latency of whatever guidance system you have. You are looking at 20-30 meters offset between frames.
Better guidance probably balloons the cost to the 10s of thousands of dollars.