Hacker Newsnew | past | comments | ask | show | jobs | submit | eggie5's commentslogin

Adyen | Founding Research Engineer, AI | ONSITE (San Francisco) | Full-time

Adyen is a global fintech platform powering payments, data, and financial services for companies like Meta, Uber, Spotify, and Nike. We’re profitable, publicly traded (AMS: ADYEN), and currently building a new AI research hub in San Francisco.

We’re hiring a Founding Research Engineer (AI) to help launch this team. You'll work directly with our SVP of Engineering and be the first AI research hire in SF, helping define our strategy and ship applied AI systems—focused on areas like foundational models, agentic workflows around identity and risk.

This is a hands-on role for someone with deep technical skill, strong product intuition, and a drive to build.

More info: https://careers.adyen.com/vacancies/6685352-founding-researc...

Questions? Reach out: matt.rum@adyen.com


Adyen | Sr DS & TL GenAI Team | Full-time | Madrid

Looking for SR DS and Team Lead roles for our tech-hub in Madrid. We have A100 GPUs and lots of use-cases were ready to execute on. Read about some of our recent work in agentic flows: https://www.adyen.com/knowledge-hub/data-agent-benchmark-for...

Looking for candidates with deep technical depth around language models.

Full relocation to Madrid.

Apply here: alex.egg[at]adyen.com and mention “HN - Who is hiring?” in the subject.


What's _online_ RL for an LLM? Saw this on the llama 3.3 reports too...


Online RL for LLMs means you are sampling from the model, scoring immediately, and passing gradients back to the model.

As opposed to, sampling from the model a bunch, getting scores offline, and then fine tuning the model on those offline scored generations.


if you have an NN that is probabilistic, how do you update the prior after sampling from the posterior?


You take the action which you computed to be optimal under the hypothetical of your posterior sample; this then yields a new observation. You add that to the dataset, and train a new NN.


ah, so observe the reward and then take a gradient step


(Well, not necessarily, which is why I framed it as training from scratch, to make it clearer that it doesn't have anything necessarily to do with SGD or HMC etc. In theory it shouldn't matter via the likelihood principle, but in practice, taking a gradient step might not give you the same model as you would if you trained from scratch. You'd like it to be a gradient step because that would save you a ton of compute, but I don't know how well Bayesian NNs actually do that. And if that works OK in supervised problems or the simplest bandit RL, it might not work in full PSRL uses because DRL is so unstable.)


just applied for GenAI position in AMS. Can you transfer visas?


to add state (memory), u can either:

* inject the running chat log into the prompt * inject the summary of the chat into the prompt


Or perhaps a progressive summary, where the most recent messages are full fidelity, and older messages get “compressed” into a summary.

You can also fine tune the model to incorporate larger amounts of data, but that may be more expensive (and slower)

This kind of sounds like human short term and long term memory. Maybe “fine tuning” is analogous what happens to our memory when we sleep.


you just explained how human memory works and i thought about implementing that in a future model that allows for more max input tokens. the further back the text, the more it goes through a "summarize this text: ..." prompt. GPT4 has 28k Token limit, so it has the brain of a cat maybe, but future models will have more max. tokens and might be able to have a human like memory that gets worse the older the memory is.

Alternatives are maybe architectures using langchain or toolformer to retrieve "memories" from a database by smart fuzzy search. But that's worse, because reasoning would only be done on that context, instead of all memories it ever had.


Years later, still puzzling over Thomas Jefferson's passivity at Philadelphia, John Adams would claim that "during the whole time I sat with him in Congress, I never heard him utter three sentence together"

Jefferson, himself would one day advise a grandson, "when I hear another express an opinion which is not mine, I say to myself, he has a right to his opinion, as I to mine." And "Why should I question it. His error does me no injury, and shall I become a Don Quixote, to bring all men by force of argument to one opinion?... Be a listener only, keep within yourself, and endeavor to establish with yourself the habit of silence, especially in politics."


> Why should I question it. His error does me no injury

Isn't that an incredibly dangerous behavior especially in political debates ? I mean, wrong political opinions do injure people. Maybe not the other people in the room but the people who elected them.

Do I miss something ?


> Do I miss something ?

Yes, the part of governmental politics that involves creating and enforcing policy. Jefferson would listen to people's arguments in legislature but infrequently debated them because his strength (in his own estimation, at least) was in writing responses, especially as policy proposals.

Someone speaking a disagreement does him no harm, even if what's being described would be harmful. But someone implementing something he viewed harmful in enforced policy is different, and he treated it very differently.

Governmental politics tends to emphasize, even glamorize, the former, but the latter is what actually affects people and Jefferson often focused his attention on it.

One could strongly suggest that Jefferson's preference for written arguments and policy over political debate and lobbying is why Hamilton had more effective Federalist influence over early policies of the United States, and manifests even more strongly in Jefferson's prescriptions for the republic — the Declaration of Independence and Bill of Rights, wanting to rewrite the Constitution every 20 years, his opposition to slavery and promotion of universal free education, freedom both of and from religion in government, strongly held national liberties and local self-government — many of which were at the time, or have always been, ignored in practice.


As Adams agreed, it was quite the perplexing philosophy. However, who won in the end? The firebrand Adams or stoic Jefferson???


Thomas Jefferson was as privileged as they come. Of course he'd have that opinion. He was at the top of the pyramid. Opinions held by his fellow man would have no impact on him personally.


yes and doesn't have traction inside ebay either :)


I don’t have inside info like you might, but it seems like a mean-spirited comment considering ebay.com uses marko, as well as other sites in the ebay group.

It definitely has not gained a lot of traction outside of ebay but there is a fair amount of buzz around the marko6 release with a completely new tags-based language.

It is also one of the few frameworks that leverage html streaming in a simple way (and has for a long time) and offers islands-style partial hydration out of the box.


everyone at ebay hates it -- the ebay blind is full of people hating on it


So you're saying it doesn't have traction internally based on personal experience or based on what you read in an app?


what do you want a signed affidavit from every ebay engineer


LOL


not a good sign. why do they hate it?


Put simply, many people are using older versions. There have been large developer experience improvements in the past year or so.


You’re explaining basic external observations to someone who is painting an internal picture of Marko’s political status within eBay. This kind of insider take is gold.


created at ebay and hated internally :)


What's the main criticism?


Speaking as a former eBay engineer.

The criticism is that it isn't React, and that nobody uses it but eBay.

There's nothing really wrong with it other than that there's no reason for it to exist when other frameworks work fine and have community traction. There's nothing special about eBay that requires it's own special UI framework.

And when we hire contractors to do anything they won't touch Marko so we have to use React anyway on a lot of projects.

Just not an efficient way to spend the companys money creating and maintaining a superfluous framework nobody asked for.


Get better contractors. It is just a JS framework. I have had to work in various weird frameworks, backend and frontend, and it is easy to get started in an existing codebase. Most of the fundamentals are the same.

I am so tired of developers who are scared of learning new things.


> And when we hire contractors to do anything they won't touch Marko

What's the point of hiring contractors if they will show zero flexibility to work with your existing stack?

A consultant could make you reevaluate your stack (and that would be the point) but IMHO contractors should just focus on getting the job done.


SEO and CWV are not good enough reasons?

Ebay is not your mom and pop e-commerce website.

If Ebay had used React the business would have suffered. That's why Amazon isn't using it either.


React is more of a library, if you want to compare Marko to something you could take Svelte, Next or Qwik. That said, React seems to be moving in the same direction with React Server Components (announced two years ago and still not available).


I think the main criticism is portability of skills, ie everyone wishes ebay would just use react.


This is why engineers shouldn't pick tools. They prioritize their resume over the quality of the product.


Most of them anyway ;)


Which in my book is a fairly weak objection. Learning a new framework does not take much time if you already know one framework and you are put to work on an existing code base. I only knew Vue when I started working on an Angular codebase and it only took me a couple of days to get up to speed.


in netherlands, for example, they (Goog) has to negotiate w/ the local-employee-backed Works Council. If they didn't bother setting one up, like Meta NL failed to, it will take about 3 months to do that. Then another 1-2 months for negotiations w/ the council. Then if they come to an agreement (they don't have to) w/ the redundancies and severance the affected can _then_ be notified.

for example, technically the meta layoffs back in November have not even happened here in NL and the affected won't even know until at least March-April!!


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

Search: