jmward01

41 ms·

811.
Executor 2.0 for MS-DOS(archive.org)
3 points by jmward01 2y ago|1 comments
812.
by jmward01 2y ago
If they can deliver the drives and support for pytorch out of the box then I know what my next laptop will have in it.
813.
by jmward01 2y ago
Getting rid of the time side then, think of it as an orbit. If the universe was expanding then something could be orbiting slower than gravity would allow. Basically this question keeps bringing me back to the ties between mass and the expa
814.
by jmward01 2y ago
I doubt anyone is still looking at this thread but I did actually start playing with RWKV by adding sacrificial training techniques to it and the results look promising, at least for early training.
815.
by jmward01 2y ago
Even a good company may not give you an equity event for??? Ever? The trend seems to be not going public and at best being acquired at some point farther out in the future. Even then the odds aren't great. There was a great HN post tha
816.
by jmward01 2y ago
These were my thoughts exactly. Even if they lost all engines/power/hydraulics they would have had 8 mins to start up the APU so gear and flaps wouldn't have been an issue and clearly they still had some control. Did they try
817.
by jmward01 2y ago
Not much technology related to this issue. Maybe we can steer the direction that way at least? How about this: Are guns more effective at quickly subduing people than other methods? Basically: Why shoot if it is less effective at stopping&#
818.
by jmward01 2y ago
The expansion of the universe has always come down to one question for me. In an expanding universe when you throw a ball up what speed does it come down?
819.
by jmward01 2y ago
> It seems to me that it's unreasonable to expect Congress itself to have informed, correct opinions on technically complex topics. Actually, that is exactly my expectation of my representatives. They have the resources and connecti
820.
by jmward01 2y ago
Yeah. I looked at the dataset and there are a lot of possible tasks you could train against here since it has some great annotations. So, having a simple reference baseline, like a pretrain gpt2 run (which I think your repo is set up to do)
821.
by jmward01 2y ago
This looks like a great dataset! Thanks for posting. I'm looking for projects just like this to try my training modifications against. Do you have initial results shown? It is a small model/dataset so training the gpt2 model in th
822.
by jmward01 2y ago
I think people are exhausted from court rulings that rarely appear to be in favor of people over companies. At a certain point the belief in an impartial, consistent and fair court system is eroded to the point of breaking. Why look/ca
823.
by jmward01 2y ago
(B)illion. It indicates the rough number of parameters in the model. Higher is generally more capable. 1B models are currently at the top end of 'easy' to deal with for playing around fine tuning and the like for most home lab set
824.
by jmward01 2y ago
I plan on checking out RWKV and seeing if I can add my sacrifical training techniques to it this weekend. There is a reason quantization works, it is because models are very badly trained right now. I think we can get really good performanc
825.
by jmward01 2y ago
The age of tiny models is just about here. We are finally busting away from the standard transformer block and training. I have a side project that can dramatically shrink networks by a set of techniques I call sacrificial training[1]. I th
826.
by jmward01 2y ago
"The optimism buoying these war industries..." What a line. I turned down a job working with good people for good pay in this industry. It will happen without me but I couldn't put my best ideas directly into creating weapons
827.
by jmward01 2y ago
Unified embeddings show an example of sacrificial training techniques, techniques that add parameters at train time that improve training performance but can be removed for prod so no impacts to prod code/weight structure.
828.
Unified Embeddings(github.com)
2 points by jmward01 2y ago|1 comments
829.
by jmward01 2y ago
This is why we need device to device encryption on top of all the security that a telco has. There is no excuse for any connection I make being unencrypted at any point except the receiver.
830.
by jmward01 2y ago
We are in post policy politics. It no-longer matters what you do, how intelligent or well executed your plans were, just what the messaging says. It is far easier to pick things that you can control messaging on and make them the issues, ev
831.
by jmward01 2y ago
Totally agree. I see nothing 'left' in the current democratic party. Pro wars, pro boarder hardening, no real effort on healthcare. Slowly but surely they are turning towards the 'conservative' path. Arguably that is one
832.
by jmward01 2y ago
He was a great human being. We need more people like him, especially in government.
833.
by jmward01 2y ago
The 4090 computer per watt is the best (on paper) between the 4060 ti, 4070 ti and 4090. Best bang for $$ though looks like the 4070ti 16GB. I've been eying that one for a new dual card training rig.
834.
by jmward01 2y ago
You can get 4060 ti 16GB cards for ~$450 or 4070 ti 16gb for ~850 instead of the $2.5k for a 4090. I wonder how well 4 of those cards would perform. The 4060 TDP is 165w instead of 450w for the 4090. The 4070 looks like the best tradeoff th
835.
by jmward01 2y ago
all my own ideas in there. I was thinking of writing it up more formally, but I am more of a 'think -> build -> next thing' kind of person.
836.
by jmward01 2y ago
The answer is simple: No matter the reason, if you have a data breach you must pay each person 100$ min with higher amounts depending on the information lost. Additionally, if that information is used in a crime then you are liable for furt
837.
by jmward01 2y ago
Current embeddings are badly trained and are massively holding back networks. A core issue is something I call 'token drag'. Low frequency tokens, when they finally come up, drag the model back towards an earlier state causing a l
838.
by jmward01 2y ago
This problem shows how antiquated military systems and thinking really are. Tracking, identifying and deciding on how to act for every thing in the sky, especially things advertising what they are from the second they take-off until they la
839.
by jmward01 2y ago
Embrace the diversity. Some people use these tools, some don't. What matters is that you build something great at the end. The only people that are 'wrong' are the ones that mock those that use the features and the ones that
840.
by jmward01 2y ago
Sounds like there are only two real strategies: - Put a lock on it (make it hard to get to. Probably by just burying it very deep and destroying access to it) - Make it boring to future people. Put household garbage on top of it. People hav

More ›