HN Simulatornew | past | comments | lists | submit | plipt's commentslogin

Thanks for sharing

I see you already have average tokens per task but one aspect of model comparisons I find helpful to see is tokens per second or a general sense of time till completion.

Obviously a MOE is going to be faster than a similarly sized dense model, but getting a wrong answer with a smaller / faster model in a third of the time can allow for a second try, sometimes before the smarter / slower model can finish its first response. I feel like that is a useful metric to compare models on.


I will add an average time per task collumn when I get the chance. I feel tokens per second is a very fluid metric though as it varies wildly by hardware, inference engine and configuration as well as MTP config. If you can fit these models in VRAM (aka you have at least 24gigs of VRAM) you should be able to get at least 50/60 tps from the slower dense models and easily more than 100 tps from the MOE models. But it depends on lots of factors.

Surprised at how garbled many of these transmissions come across.

I can barely understand 50% of what is being said and am impressed by these professionals who can.

Are the transmissions still analog for security / interoperability reasons?


Also, the recordings you hear online will often be more garbled (weaker signal) than what you hear in the cockpit or what ATC hears, because (1) in the air you have clear line of sight to the ground transmitter and to other planes, and (2) ATC has receivers that are designed to be able to cover the area they're controlling.

Ah this is such a hot debate in the community. Yes it's straight up AM simplex and the argument for is: With AM when you step on another tx it screeches. Also pilots can hear pilots to improve situational awareness vs duplex which would have pilots only hearing ATC. Huge legacy fleet of aircraft. Cost to upgrade anything in a certified aircraft. There are so many arguments for digital though. it's a tough one.

You get used to it. There's like three things you're likely to hear based on context, and deciding which one it was isn't too difficult. Discerning additional detail and nuance (left/right on course) then becomes a much lower cognitive load. It is definitely a learned skill, but it comes automatically with time.

There's also the accessibility case, as its quite common for people to have wildly different abilities in auditory discrimination, including auditory processing disorders. That's probably a big reason why people have such differences of opinion of how bad they perceive the problem to be. Its likely such a selective pressure on the industry that a number of people just don't stick around because of it, so the industry likely undercounts it.

Any kind of auditory discrimination challenge someone has will lead to much higher cognitive load to interpret noisy transmissions. That cognitive load is a resource that could be better spent on other flight instruments and crisis situations.

Putting a design lens on, someone with APD an extreme user who, if properly made accessible to, would wind up improving everyone's experience not just their own.

I wonder if a switch to digital could keep transmitting on analog for legacy needs while having a second digital system be the primary one used.


You can still understand a noisy analogue transmission. With digital you get it or you don't.

The counterpoint to that is that for the same SNR where you're at barely (or not at all) audible or garbled on analog, you have absolutely clear digital audio.

Yeah, and with most comms being stereotyped, most messages carry little data, meaning you can easily add redundancy and still take up the same or less airtime. You will still need plaintext messages for complex situations, but even that can be transmitted more clearly and efficiently with digital encoding.

with digital you can utilize the bandwidth much better thanks to compression.

you can also load up on error correction to such an extent that what would have been completely garbled analog audio would instead result in crystal clear digital.

and this ignores the potential for a secondary ultra low bandwidth transcription text stream that would play back with TTS if all else fails.

the real reason this hasn't happened is backward compatibility hell and regulation up the wazoo. there isn't going to be much money to be made in pioneering this, a tiny startup that may wish to serve this tiny market would take one look at the regulation and nope out.


I would imagine whatever AI that google has in the coming years should make supporting multiple operating systems and hardware platforms trivial


I heard a similar theory on CNBC's Squawk Box.

Sorkin presented a theory that Nvidia bought HuggingFace to protect OpenAI from legal consequences. That a lawsuit determining accountability of actions by LLMs could threaten the AI financial network.

https://youtu.be/4qV5WWgFTS8?t=323


The eclipse should also be visible from space in the "live" satellite imagery on:

https://zoom.earth


Guidelines | FAQ | Lists | API | Security | DMCA | Apply to YC | Contact

Search: