Eggpants
Eggpants on Hacker News
I don’t understand what “floating-point math is non-associative” has to do with the random selection of possible result token set after the soft max operation. The weights of the model are set and don’t change during the many matrix operations(which is deterministic regardless of the floating point encoding scheme).
Honest question about the non-deterministic nature of llms. Every programming language provides some way to control the random seed. Is this not true in GPU programming land?
I was stationed on the Lincoln during the first gulf war.
Most folks have 12 hour shifts, you wake up in a triple stacked “coffin” bunk(I couldn’t roll over without my shoulders hitting the top of the bunk) and wait in line to shower, then wait in line for each meal all while performing your responsibilities. Plus repeated fire drills in full gear because if a fire isn’t put out folks die. I certainly remember losing all the scheduled liberty ports due to the war. It definitely builds up your stress levels over time to breaking point levels. I didn’t see anyone attempting to jump overboard but did see some purposely hurt themselves.
The only port large enough for a carrier (at the time) was/is Dubai but that just makes the ship a sitting duck right now. I can’t imagine going that long without a break.
Of course it’s never mentioned that all these hacking techniques are in the training set…
Would you also consider a database of questions and answers smart? LLM are basically lossy text compression databases with a clever query method. Useful for sure but it’s not thinking, it’s recall.
Just look at some training sets to see how the sausage is made: https://huggingface.co/datasets/nickrosh/Evol-Instruct-Code-...
Talk about a giant whiff.
I was hoping they would announce their first open weights model, perhaps an older model they don’t offer anymore, but no. Instead he get this bs statement that reeks of “dam it I’m so close to being a billionaire” desperation. Not even acknowledgement of how much data they stole from others yet he whines about distilling.
It’s like his goal in life is to be a Scooby-Doo villain.
And yet two orders of magnitude from making a profit…
I've been pretty happy with LLM svg based data plots I've asked, including log scaled axises and histograms. Definitely a first world problem of course.
The construction of data centers brings jobs, most of the running of data centers is done remotely. There are more security guards than on site tech workers. Just look at the size of the parking lots. In Ashburn Va, a single tech will be assigned to multiple data centers further reducing the number of “decent jobs”.
Then it would be fair game to have an AI answer the hiring AI questions. After all, they now require the use of AI on the job right?
“annualized revenue run rate” is a bogus accounting term. It’s like taking a paycheck and multiplying it by 365. Notice the complete lack of any mention of profits.
I asked it to write a python program to find the intersection of a line and a circle and it one shot it. Asking for a Swift version worked as well.
Rust version goofed with lines of comments and code intertwined.
I asked what highway connects two given cities that it got wrong but would have been correct if I asked for two other cities.
I asked it to save answers to a file or notes but no burrito
Asked for an ascii drawing of a race car. Not sure what that thing was it returned.
I’m old enough to remember when one checked the assembly a compiler generated because early on they produced terrible assembly. Eventually they got good enough to not needing to check.
Coding agents are not close to that yet, but it’s interesting watching history repeat itself.
This narrative of the coding agents being so much better now over the last few months seems VERY exaggerated. I’m still spending a lot of time telling Claude: No, that didn’t fix the problem. Again. Can you handle this task or do I have to give it to codex?
What I found interesting is nanoclaw isn’t a working product out of the box. You must use a coding agent to complete it with features you want. For example add iMessage support, etc.
In other words, Claude is the compiler.
I’m using this but using gpt-oss-120B instead of a cloud service. It has been eye opening when I realized the LLM is beings used as a compiler.
I asked it to add apple iMessage and apple notes support as I I rather have long responses, like write me a program ideas, not fill my iMessage history. The local LLM, which I believe has limited bash training data, does pretty well.
For example: I enjoy industrial music and asked it for the tour data of the band KMFDM which returned they will be in Las Vegas in April for a festival(Sick new world). This festival has something like 20 bands most of which I never heard of. I asked nanoclaw to search all of the band list and generate a listing grouped by the type of music they play: Industrial, rap, etc. It did a good job based on bands I do know.
I was pleased as I certainly did not want to do 20 band web searches by hand. It’s still at a bar trick level.
It gives me hope that an upgraded agent based Siri-like OS component could actually be useful from time to time.
What’s interesting is Anthropic being singled out here. That either means:
1- OpenAI, Microsoft, Google, Amazon, etc have no problem with their products being used to kill people so no need to bully them.
2- These other products are so terrible at the task that the clown shoe wearing SecDef is forced to try to bully Anthropic.
I read his book “A new kind of science” and quickly figured out why it was self-published. My goodness it’s bad and need of an editor.
A big disappointment as I’m a fan of his technical work.
Last night, I was able to modify nanoclaw, which runs in a container, to use iMessage(instead of whatsapp ) and use GPT-OSS-120B(instead of Claude) hosted on a Nvidia spark running llama.cpp.
It works but a bit slow when asking for web based info. Took a couple of minutes to return a stock price closing value. Trying it again this morning returned an answer in a couple of seconds so perhaps that was just a network blip.
It did get confused when scheduling times as the UTC date time was past midnight but my local EST time was before midnight. This caused my test case case of “tomorrow morning at 7am send me the current Olympic county medal count” test to be scheduled a day later. I told it to assume EST timezone and it appeared to work when translating times but not dates.
After Claude failing to fix a bug after 4 attempts, I wrote this prompt and sure enough the bug was fixed. lol.
> no it still doesn't work. Do I have to let OpenAI fix this or can you handle it?
I can definitely handle this. Let me try a completely different approach - …
Back when I worked at AWS, I wrote a one page “press release” about this concept where the value add was the weight/power reduction of customer sats that could offload on board processing to the “Cloud above the Clouds”. There is a bottleneck with the down/uplink ground stations to do the on the ground processing.
The biggest problem folks had was even with equipment with 99.9% reliability something breaks every day due the huge raw number of devices involved. And most network equipment is not any where close to being radiation hardened.
I had some fun with it with Bezo’s fist bumping folks because SpaceX was cleaning BlueOrigin clock.
I talked to one of their lawyers and didn’t hear anything afterwards. I left AWS and a couple of years later Amazon announced AWS ground station. I wonder how much my paper contributed to green lighting that project.
Definitely helped that I just missed the punch card era. I know young me would have dropped my stack of cards many times…
Young me remembers fondly poking and peeking system memory locations to see what happens. The manual, if I remember right, had a table of memory locations to system settings. Things like font and background colors.
I made a “punch out like” boxing game in basic where the background color blocks was the opponent and the font lines was your character via poking memory locations.
It was slow but I was just a kid at the time. It definitely told me what I wanted to do for a living at an early age.
And yet Tesla won’t take responsibility for FSD mistakes. I had one and it’s amazing when it works but it did try to kill me a number of times.
OpenAI and Microsoft have defined AGI as a revenue number so yeah maybe using that definition.
I believe AGI will require the ability to self tune its own Neutral network coefficients which the current tech cannot do because I can’t deduce it’s own errors. Oh sorry “hallucinations”. Developing brains learn from both pain and verbal feedback (no, not food!) etc.
It’s an interesting problem where just telling a LLM model it’s wrong is not enough to adjust Billions of parameters with.
I’ve been doing the same with GPT-OSS-120B and have been impressed.
Only gotcha is Claude code expects a 200k context window while that model max supports 130k or so. I have to do a /compress when it gets close. I’ll have to see if there is a way to set the max context window in CC.
Been pretty happy with the results so far as long as I keep the tasks small and self contained.
The M-series unified memory is built into the chip itself, not separate components. Of course Apple is going to maintain their margins, but it’s easy to see why with this design more memory is more expensive than drams. Well maybe not with the current market pricing which hopefully is temporary.
Because it was really about serving red meat to the MAGA base. My relatives in Kentucky cheered because they believed the “all those lazy blacks are getting fired” narrative.
It’s so strange to me as they are deeply religious and some of the kindest folks I know, but also the most racist.
I agree that’s Zed is very nice but a ticking enshitiation time bomb.
However, you do see how this autocomplete “feature” and its whole copilot everywhere strategy, is about M$’s roi of its ai investment right?
So if you have a problem with VC money you should stop using VSCode as well.
Of course they will just invent more accounting terms, like they do with azure, to hide how much money they are losing on it.
https://windowsreport.com/steve-ballmer-calls-bullshit-micro...