blackkettle

I’m not disputing the reality; it would be a bit ridiculous and I assume incredibly dangerous especially given the rate at which the whole thing is moving. I’m just arguing that the article is worded in a way that is encourages such a misunderstanding.

The article definitely makes it sound like they had “boots on the ground”. There’s a picture of muddy boots and the caption says they are covered in ancient mud.

This is pretty much exactly how the public education system is structured here in Switzerland. Early years focus primarily on socialization, multilingualism, literacy and basic numeracy. Then as you move through later elementary school to middle and finally through the Matura phase there are rapid accelerations and bifurcations. If you can keep up you continue on the path to ETHZ and other top Unis. If not you move into the trade or technical school or lower Uni tracks. There is a chance to come back but the baseline expectatio. Is that about 20% of students will go straight through the Matura process. If you complete you have basically free access to any university in Switzerland. ETHZ is something like CHF800/semester. But the trade and technical school path is also good IMO. It’s appropriate.

It’s also not a perfect system but much much better than what I grew up with in the US compared to what I’m seeing now with my son here.

Comparison by competitor but it’s believable IMO. Basically about the same performance as whisper:

- https://deepgram.com/learn/nova-speech-to-text-whisper-api

Not surprising though as at this level all these options are starting to be leveled by inconsistencies in manual groundtruth. Conformer alone also isn’t the most powerful architecture out there for speech. This is also slower than, say running a large k2 zipformer via onnx on cpu.

Also if you have a small shop at this point you can do all of this yourself with whisper large v2 on a single 16gb gpu via some tweaking of https://github.com/guillaumekln/faster-whisper and an OSS LLM.

Interesting stuff but I think margins in this space are getting ready to simply vanish.

Have you seen the recent work at TinyStories: - https://arxiv.org/abs/2305.07759

It got some nice attention here: - https://github.com/karpathy/llama2.c

I think there may be some applications in this limited space that are worth looking into. You won’t replicate GPT-anything but it may be possible to solve some nice problems very much more efficiently that one would expect at first.

I’ve started doing this with ASR hypotheses from colloquial spontaneous speech. It tends to have similar issues. Lots of shady human ground truth especially where addresses, alphanumeric sequences, repairs and repetitions and other essentially non read speech are concerned. The very large Whisper models are consistent in their transcription style and highly reliable as long as you pick strongly represented languages. And ChatGPT can do a very good job at comparing the linguistic coherence of hypotheses from multiple recognizers. Together these models can annotate, analyze and ingest far more data more consistently than human annotators at this point (at least in the best covered languages). We haven’t quite realized this as a community yet though, because the standard datasets we use for evaluation contain all these human inconsistencies. Wild times.

Eat less, move more, make sure you sleep.

I also found it odd looking through the whole piece waiting for an example only to reach the end with no idea what these daydreams actually are?

Article mentions tubi. This is also available and included in prime - agree it’s an amazing piece.

It probably means they have tried it for _some_ purpose, but not necessarily the one described in OP's post here. The claim is that this is specifically useful for quantization. It's seems reasonable to assume that this would have initially been tried and potentially discarded for having little or impact on general accuracy. But that's a different issue. I suppose we'll here something definitive in a month or so.

Most STT systems also tend to still train on normalized text which is free of the punctuation and capitalization complexities and other content you find in text LLMs. I suspect we continue in this way in part due to lack of large scale resources for training, and due to quality issues - Whisper being an outlier here. Anecdotally 8bit quantization of larger pre-normalized STT models seems to not suffer the same degradation you see with LLMs but I can't speak to whether that's due to this issue.

Just a random guess, but I would hazard that it's not the heat. Going in a properly hot sauna and 'soaking' for 15min or more gets you really sweating, and increases your heart rate. Basically it 'gets your blood pumping', and _probably_ produces an activity profile not too different from mild exercise; and regular exercise has also been shown to consistently lower risk.

I'm curious as to why you think a linter tool for sanitizing the text is less effective than the other measures you describe. As long as the target problem is zero-width characters, this seems like it should be more effective than all but #1. I would agree if you include the synonym or other lexical fingerprints.

Moreover, the whole 'smart bathroom' thing has already been pretty much solved in Japan - and more or less keeping in mind the principle that you describe. You can get your temperature, get a warm seat, turn on some ambient music to mask your own, more personal music, flush, bidet, whatever all in most toilets in your average department store.

You can get more complex offerings that do perform some more complicated tests if you really want to.

The 'bath room' itself is also typically a single, self contained unit, often in some kind of mold, and including the shower, accoutrements, and bath in a full space designed to get wet and save/share the clean water in the bath.

The 'bath room' usually has a waterproof, electronic control panel which allows you to set timers for filling the bath, set an ambient temperature to keep it at, set the preferred level of water, set different default temperatures for the shower and bath, and typically also access all these features from other places like the kitchen where a companion panel is installed. This is typical of your average house - I just described my wife's parents' place, which was built 10 years ago.

The whole design thing is completely solved there, and why that approach has not been adopted across the U.S. and Europe by now is completely beyond me.

But the killer thing there would be to bring the Japanese approach to the US and Europe and plug in some better smart access features. Or go to Japan and plug them in there, where the infrastructure and demand is already in place [well, Google is probably on it already but...]

According to wikipedia the road from R&D conception to commercial deployment with Singapore Airlines for the A380 spanned almost 20 years, from 1988 to 2007.

How many software projects have that sort of scope?

They've set themselves up perfectly for their next product/service:

ResumeDetected: detect resume plagiarism; our solution goes beyond verbatim phrase matching and identifies risky and dishonest behavior using our state-of-the-art proprietary deep-learning systems and one-of-a-kind annotated corpus of real world resumes.

Pay to get the clever phrases. Pay to see who paid to get the clever phrases. Pay more to obfuscate your data from those who paid to see if you paid to get them!

I love how they describe it as a 'scam', but this is exactly what corporations do all the time. He received a salary, and in return completed his work in a timely, efficient and apparently excellent fashion.

The 'an' rule for 'historic' is related to historical pronunciation trends. Particularly in Britain, but even in the US if you go back a bit further, it used to be common in certain dialects to not pronounce the 'h'. The choice of 'a' vs 'an' reflected that trend:

http://www.betterwritingskills.com/tip-w005.html

I think it has more to do with the fact that they are finally starting to allow their researchers to publish. They were platinum sponsors at INTERSPEECH 2017 this week, and actually published a paper there. I'm pretty sure that was the first time _ever_ despite their recruiters showing up every year.

There area a number of comments in this thread similar to yours on the topic of university as an 'experience'. I started my undergrad degree in 1999, and I think that was very much the case at that time as well. It was the same for the dorm community that is mentioned as another reason in other comments.

All of that was pretty similar then, and this was when UC tuition was still very, very reasonable.

The main difference, maybe, was that at that time the experience was less controlled, and less managed. The dorm I lived in first year was run by fourth year undergrads and grad students, and their responsibility was basically to make sure that no one was opening vodka bottles in the common room, or setting off smoke detectors with bongs or puking in the elevator. They were compensated with free or heavily reduced room and board.

The experience was less controlled and less micro managed across all aspects of university life at that time, and even more so prior. My admittedly unsubstantiated viewpoint is that the proliferation of control mechanisms, and transition of traditionally subsidized experience management to formal salaried roles is a big contributor to the unnecessary increase in bureaucracy, and it has a negative long term effect on learning from both the academic and the 'experience' side of things.

> On the other hand, if you’re working in an agency/consultancy where it’s important for security/reliability/maintainability reasons to keep up to date with various areas of front end tooling, database types, AI, serverless, etc, then the answer veers closer to ‘never’.

Well I mean, the fact that they mention AI in the same breath and as though it entails the same scope as frontend tooling seems a bit absurd. People with PhDs in 'AI' spend much of their professional lives keeping up-to-date with a very narrow subfield: "keyword-spotting for automatic speech recognition".

The idea that a 'full stack' developer somehow even encompasses that certainly takes it to the realm of complete fantasy.

But "competent to be usefully productive across the development spectrum", is surely possible.

My go to resource for this issue is:

* http://www.regular-expressions.info/tools.html

nooks and crannies of pretty much every major flavor and variant, prepared as an easy reference.

Exactly, and the simple, physically decentralized, distributed, and somewhat resource/time intensive aspects of paper-based balloting are all great checks against easy attacks.

The fact that virtually anyone both capable and eligible to vote is also capable of understanding the voting process, as well as what kinds of physical acts are signs of fraud means that many more people can evaluate the process, and determine - even just by looking - whether something fishy is going on at their voting station.

Your average voter will be easily convinced that he "doesn't know enough to judge" whether something is fishy with his voting machine, even if something seems to be clearly malfunctioning; "oh don't worry sir, the print out might be wrong, but your vote was definitely counted correctly internally".

The rise of any form of electronic voting is really troubling - and don't even get started on anonymity concerns...

No. To protect voting, don't use software. Everyone needs to be able to _understand_ as well as be able to verify that they successfully voted.

Besides the issues with what software the machine is actually running, most people cannot comprehend or understand that software - even if it is open source. That is not acceptable for an open democratic society, or to sustaining it.

In this particular situation it should not be necessary to rely on an expert to explain whether the vote counting mechanism is reliable. This only adds to the problem of unreliable or scheming officials - it doesn't improve anything in terms of transparency.

I would assume they are only going to be doing the raw data collection and maybe cleanup and annotation, and that data should be made available so you can train what you like.

If you poke around github and the Kaldi lists a bit more you can see that they are experimenting with and probably planning to use Kaldi.

I wonder what they plan to do for provisioning. It is one thing to collect data and train models, but quite another to make the service available over the web in an unlimited capacity. And we are not yet to the point where you can reasonably expect to run a high quality open-vocabulary STT system in your browser. The search network is typically in the GBs range.

there are no tick marks on the x axis.

and this may be even more problematic that it appears.

the ones from category one might have difficulty knowing what to do - unless we're talking about the lyft CEO they know success from something other than uber. that might not necessarily be good.

the ones from category two might have the same trouble as one, but add to that that they will have to compete/convince the board of probably every little move.

and if they don't do something fairly dramatic, fairly quickly, the good people at the mid and lower levels will start to flee for greener pastures.

If you don't understand something and/or cannot explain or describe it, should you believe it? I think this is a very difficult question.

Wait, what? Is that what we were expecting - or even hoping for?

> In adulthood, many prodigies become experts in their fields and leaders in their organizations. Yet “only a fraction of gifted children eventually become revolutionary adult creators,” laments the psychologist Ellen Winner.

Why on earth is that something to 'lament'?