I've been deeply impressed with Jev as it made a bunch of workloads we had on Luna or Gemini 10x cheaper and 2x faster (previously used non reasoning version for latency reasons).
Now Laya promises another speed up and it's open source. Tbh if it can't run on a CPU I anyway want to buy it from an inference provider. Managing gpus in production is a non trivial problem.
What I also wondered about Jev is how different it is from something like tabular foundation models. They seem to overlap in use cases. Which then leads to the question, what is actually learned? A lot of people in machine learning spend time to making things explainable and always struggled to move beyond data induced biases.
Having it open source is awesome as fine tuning might give additional performance on the task we care about.
I have 10000+ inventory items to categorize but I need an intelligent model (not just if statements). Using LLMs has been slow and expensive and I needed to queue it to run for hours. Jev did it in minutes and for less than 1 cent
The same level of intelligence gets roughly 10x cheaper per year. So you might both be correct where a large part are commodity tasks but frontier is hard and valuable and not commodities.
But that's supply and demand, not technology. Right now a lot more people want their inference than they can supply. as supply catches up in the next 5-10 years, the underlying tech at scale is probably cheaper than GPUs per token produced.
This is obviously cool and useful so kudos, but wow security researchers have to throw their hands up all the time.
Now we open another attack surface where you can ask a remote agent to do things by default. There was a time when you call this a Remote Code Execution vuln. It's of course a feature here.
Remote Control requires (relatively) manual pairing per session and setting the controlling session as trusted, even if the controlling session is a web browser. You can enable a more well-encompassing remote control "server" mode, but it requires extra configuration that isn't (at least clearly) visible in the app itself.
Pretty much nothing with minimum viable guardrails (restricted shell into Container into MicroVM into VM + Firewall) which any dev knows how to setup, the fact that most don't is mostly about lazyness/recklessness.
Which is why the real answer to this is actually an older paradigm: apprenticeship.
Companies should be doing their own training for their jobs, and university can be freed up to actually be about intellectual pursuit rather than economic pursuit.
It was companies themselves who created this issue anyways, by demanding degrees for every damn job, regardless of whether a degree could prepare you in the first place, and slashing entry level jobs to almost nothing. Then they started using degrees as a proxy for work experience (and I mean that in the literal sense; "bachelor or 4 years of experience" is common to see in job reqs, despite no bachelor degree actually being equivalent to 4 years of full-time job experience), which just further pushed incoming students to expect that to be the purpose of their time at University.
> Companies should be doing their own training for their jobs, and university can be freed up to actually be about intellectual pursuit rather than economic pursuit.
This would be great of course, but the reason that 'what actually happens' is so different than 'what should happen' are the usual economic incentives. How can we change the financial calculations to separate formal education from job training in a way that drives the ship in the better direction.
One way could be to regulate degree requirements to only jobs that really require them like engineering, nursing, etc. Another might be allow student loan bankruptcy if certain criteria are met such such as x number of years below y income.
We (meaning the US) won't ever fix this ourselves, because we systematically refuse to hold businesses to any kind of standards.
Make it illegal to discriminate based on degree status in hiring (it's not like anyone is hiring an SWE without doing technical interviews anyways, so if you prove out their skills why does the paper matter?), and you remove both the backdoor socioeconomic discrimination route (which is itself a backdoor route to discriminate against protected attributes), and solve the degree 'inflation' incentive.
Yes, but that's not something universities can solve.
Universities predate capitalism (despite the author's absurd assertion that they were founded to serve the interests of businesses), and will exist after it, if any humans survive it.
It does scale even better than anything else. Majority of students are just not very good at it. And proving that they have learned sufficiently is hard. Thus schools and universities.
I sense a certain contradiction? If majority of students are just not very good at self-studying methods, then they don't scale well.
Mediocre student that sits through lectures, does the homework and submits assignment on due dates will get a good level of mastery in the subject. The very same student would be totally lost relying only on self-studying - he would not even know what he doesn't know.
It is in sense contradiction. Schools and universities are greatly limited by facilities and teachers. Self studying is not. Thus with modern technology or even with just mass printin you could theoretically educate as many students as you want. In that sense I do say it scales much better.
I fully admit that for efficient time spend and production of sufficiently educated students it is not effective method.
Well, we can't have the counterfactual world but could it be that without computers, Internet or AI, there might be no/less growth?
I somehow doubt that AI will accelerate growth constantly as the world has inertia and I'm not sure AI is sustainably increasing our ability to change. On the other side, if there is a technology that can do that it could be intelligence.
Now Laya promises another speed up and it's open source. Tbh if it can't run on a CPU I anyway want to buy it from an inference provider. Managing gpus in production is a non trivial problem.
What I also wondered about Jev is how different it is from something like tabular foundation models. They seem to overlap in use cases. Which then leads to the question, what is actually learned? A lot of people in machine learning spend time to making things explainable and always struggled to move beyond data induced biases.
Having it open source is awesome as fine tuning might give additional performance on the task we care about.
reply