sfriedr

Location: European Union

Remote: Yes

Willing to relocate: Inside EU

Technologies: LLMs -> Python

Résumé/CV: On request

Email: [email protected]

........................

I have two distinct graduate degrees, one in CS and one in pure mathematics, before I started with AI/ML.

I can work part-time (either regularly a few days per week or larger but fixed blocks of time).

Tell me what your needs are, so we can see if we can make it.

> Unless they have been shown to be relevant in any way, I don't see why the paper need cite them.

Fair. Your argument then falls precisely into category (C) of the four mutually exclusive options I outlined above.

But you'd then need to argue why the 6 models you compared against is the comprehensive model sample to test against, that contains -- and not just some arbitrary set of recent models that happen to be dominated by the newly proposed model. (And maybe that is indeed the case; then it should be easy enough to update the arxiv draft by incorporating a section where you argue along those lines.)

Not, but having a small section in the paper would be reasonable, that illustrates why the most pertinent models that might be relevant at first sight (like the ones I cited) are actually not applicable.

The onus is on the authors to place their research in context and provide compelling arguments - not on the reader to guess why their model was compared against model A, but not model B.

What do I mean by "pertinent"? Of course it is not necessary to cite every "All You Need" paper.

But:

(A) I'd argue it would be necessary to cite those "All You Need" papers that have either gathered a fair amount of citations or media attention (which is the case for both of the papers I linked to), or are meaningful descendants (in the "has been cited by" tree) of thise papers. As I said, this is not really my field - but I would say there is some change that among the hundreds of papers that have cited the papers I linked, some have been scaled up to LLM levels and use basically the same MetaFormer/Hopfield architecture.

(B) If the above isn't the case and none of those models have been scaled to LLM levels - that's fine too. But then please tell the readers that you did due diligence and found that there actually is this gap in the literature (of course, feel free to close it yourself and then be the first one to train one of those models to such scales; that's the reward for doing a solid literature review - and who knows, maybe you stumble upon an even better model that will get you many citations).

(C) If you cannot perform a comprehensive literature search, but the models you compare against cover 90% of the models that out there (in production or research), and you can back that claim up - then, of course, you're safe too and I'd be very happy to be able concongratulate you that you really dis manage to achieve a breakthrough.

(D) Even if none of these things applt and it just costs too much in terms of computational power to train many other potentially competing models, or is too cumbersome to carry out a a comprehensive literature reading- that's also fine. You can then simply constrain your paper more and consider a more precisely defines slice of models, for which you then can actually do a through literature review and comparison. Then you'd also need to adapt your title though, so that it reflects the more precise scope. And please don't take this negatively, as a reader I'd much rather have a model that is proven with the highest scientific standard to be state-of-the-art on a more narrow scope, rather than a more broad claim, with only a moderate amount of evidence backing it up.

So, PLEASE, don't leave us guessing!

If you have a new candidate model that you claim is the "successor" -a strong word- but compare it to just 6 other models, and importantly, you don't let the reader know which of the options (A) to (C) apply, then you have to go with (D).

Machine learning is already too full with papers whose titles are overly broad. Somehow, other scientific disciplines have a much more sober title formulations, yet ML insists on colorful titles that usually are not particularly informative (and yes, "XYZ is All You Need" I consider to be an example of such a bad title).

I only spent a few minutes skimming thr paper, but:

1) there are a lot of papers claiming to be the successor to the Transformer, and not all of them are cited; e.g., the MetaFormer is missing https://arxiv.org/abs/2111.11418. Another candidate that wasn't compares against (or at least argued why it wouldn't make sense to compare against) are the Hopfield Networks https://arxiv.org/abs/2008.02217.

So until a more solid Related Work section is written (their section is actually called "Relation to and Differences from Previous Methods") I reserve the right to be skeptical whether their model is the "best" successor to the Transformer.

2) they say in the abstract "We theoretically derive the connection between recurrence and attention" but I couldn't find a longer theorem-proof section. So either this is done only in a cursory manner, or the proof is very easy. Recurrence and attention have been around for a long time as concepts, so surely there are already proofs in similar contexts of this fact (I am not working in this particular area of Machine Learning, so I don't know the SOTA by heart, but I strongly suspect that these aspects have been discusses previously; thr Hopfield Network paper I linked to unearthes some theoretical facts about attention, for example).

So -based on my very cursory reading- this paper seems like an interesting approach, but I do see some holes in thr execution. Time will tell whether Rentetive Network will become mainstream or not.

Ok, this was my five minute review of the paper. Now I have to urgently return to completing my actual reviews for NeurIPS, haha.

Of this is true, it would be something close of an insane situation: One of the largest datasets, that the largest companies are using to train their models (probably; many of the best LLMs have technical reports that raise more questions rather than answer them) being forced to live an obscure existance on torrents.

From a scientific point of view this is very problematic because few safeguards exist that guarantee that the dataset is not tampered with (as is the case if you'd upload it to Zenodo, which providea some guarantee of immutability).

How about trying to upload the Pile to Zenodo? Only half-joking :D

Could you share more about copyright? For example, aren't you worried that now, with all kinds of lawsuits happening [1] and copyright issues that were found in existing datasets [2], that you might get threatening letters from a lawyer some day?

I'm the author of [3] where we introduced one of the first natural-language datasets that test graduate mathematics for LLMs, but some of the prompts we took from a copyrighted book and therefore thought about excluding them. Having them in the public dataset would be really nice though, hence I'm keen about your experience.

I'd also be keen to hear how your challenge against the DMCA on sharing LLaMA's weights goes?

[1] https://www.theguardian.com/books/2023/jul/05/authors-file-a... [2] https://arxiv.org/abs/2105.05241 [3] https://arxiv.org/abs/2301.13867

Congratulation, great paper! It should have been put on HN earlier ;)

I have a few questions:

* you say (page 4): "We then perform standard instruction finetuning on the base LLaMA-7B model" Could you perhaps provide a reference to the _exact_ finetuning approach you used? I'm afraid different groups of people have a different notion of "standart" (see for example pages 131-155 from https://arxiv.org/abs/2302.08575 for various fine-tuning approaches) and without knowing exactly how fine-tuning was carried out, it can be very difficult reproduce your research and results exactly.

* the idea of using AST Sub-Tree Matching is nice. Could you please let me know which function in which file from your GitHub repository this is implemented in?

Again, great job on publishing this paper!

---

Best regards,

friederrr.org

In February I published a paper on mathematics + ChatGPT (https://arxiv.org/abs/2301.13867) and my colleagues, who were also on LinkedIn, told me it garnered quite a bit of attention.

So I signed up to LinkedIn on 3rd Feb. 2023, thinking that now would be a good time to step out of the academic ivory tower.

The thing is that I used my Gmail account to sign into LinkedIn. Which worked fine and I was a happy user for about a month. I then went travelling and when I returned one month later to log in again with Gmail - suprise! A new account was created on the spot.

So now I have the old account which had started to accumulate some important contacts: https://ro.linkedin.com/in/simon-frieder-5b6874264?trk=publi...

as well as a new, useless account:

https://www.linkedin.com/in/simon-frieder-b7901b271/

I had been happy to use my Gmail account to login, so I did not bother to setup an extra email login process into LinkedIn.

How can it be that the software engineering of a multi-million company even allows such blunders to happen?

Location: Oxford, UK Remote: Preferred

Willing to relocate: Depending on location (inside EU preferred, but willing to consider other options depending on the work)

Technologies: LLMs, other Deep Learning models (e.g. Vision Transformer), prompt engineering, PyTorch, LLMs, mathematical analysis

Résumé/CV: On request (recent PhD in ML)

Email: [email protected] Website: friederrr.org

........................

I offer machine/deep learning consultancy: happy to help you understand the state-of-art research and translate models implemented in recent papers into production code.

I can only work part-time (either regularly a few days per week or larger but fixed blocks of time), because I have ongoing research projects in machine learning that I need to finish, though I can work on these remotely if I need to relocate.

I have two distinct graduate degrees, one in CS and one in pure mathematics and an eye for detail, which is crucial when it comes to doing a sensible implementation of certain neural networks.

I also have expertise managing groups of data scientists and reseachers to pull of bigger scientific projects. My latest one was about about engineering and LLMs about which I'm particularly excited: https://arxiv.org/abs/2301.13867. It has been featured in the media, e.g. https://syncedreview.com/2023/02/03/genius-or-subpar-ai-math....

> OpenAI's chosen not to release any real details about GPT-4

Actually, they have release some details about it, in this 99-page technical report https://arxiv.org/abs/2303.08774 (which is actually two papers stitches together, once you read it; oddly enough using different fonts).

But I'm not sure if this content qualifies as "real details".

Indeed!

This is the weirdest thing. How can it be that it depends on whether I click or copy to the browser whether it works?...

Location: Oxford, UK

Remote: Preferred

Willing to relocate: Depending on location (inside EU preferred, but willing to consider other options depending on the work)

Technologies: LLMs, other Deep Learning models (e.g. Vision Transformer), prompt engineering, PyTorch, LLMs, mathematical analysis

Résumé/CV: On request (recent PhD in ML)

Email: [email protected] Website: friederrr.org

........................

I offer machine/deep learning consultancy: happy to help you understand the state-of-art research and translate models implemented in recent papers into production code.

I can only work part-time (either regularly a few days per week or larger but fixed blocks of time), because I have ongoing research projects in machine learning that I need to finish, though I can work on these remotely if I need to relocate.

I have two distinct graduate degrees, one in CS and one in pure mathematics and an eye for detail, which is crucial when it comes to doing a sensible implementation of certain neural networks.

I also have expertise managing groups of data scientists and reseachers to pull of bigger scientific projects. My latest one was about about engineering and LLMs about which I'm particularly excited: https://arxiv.org/abs/2301.13867. It has been featured in the media, e.g. https://syncedreview.com/2023/02/03/genius-or-subpar-ai-math....

> I thought of reaching for Mathematica but resorted to pencil and paper, like a barbarian.

A very computer-science perspective with the obligatory dose of hubris.

Pencil and paper is continuing to serve a much older, much more consolidated discipline (mathematics) well for... ever since paper was invented.

A lot of facets of this power balance have been mentioned.

Here's a new one (I think):

Imagine busy CEO B that doesn't care about your time, that you need to pester with emails to get his attention.

Now imagine thinking-along CEO T that tries to have a well organized inbox, possibly assisted by a secretary, so that when you send him an email you receive an auto-response telling you when and if you should get a response and how you should interpret a no-response; this may include letting people know that too complicated or unclear messages will go unresponded. (Richard Stallmann had this approach when I mailed him years ago, before his scandal, where I'd get a response with an average waiting time in hours, pretty neat. Using NLP probably anyone could filter his/her mail quite efficiently these days.)

Who would you want to have as a boss?

Yes, it's a power balance. But all parties should wield their power responsibly.

As we speak nature is serving us with a particulay nasty "decision" called climate change.

No, we're still at the mercy but just on a different scale. Earlier it was the individual, now it is the population.

I do wonder what it was specifically that you didn't like training NN's. At least when doing research on NN's, everything is very interesting, as many many aspects of them aren't well understood.

That's not quite true, as there are overlaps. To name just one big topic: Pseudorandom number generators. Here you have a (number) theory, including things like finite fields, to generate and understand deterministic numeric sequences, but also a lot of statistical methods tonassess whether these "look" random. Knuth has an entire chspter on these.

The free/proprietary distinction is only irrelevant for small-ish projects, on the order of a few months, where "good" is a one-dimensional measure in terms of features the application has: The risk of the software maintaining company being bankrupt in this timespan is low. A typical mid-level employee at a company might be very happy with such tools, especially because they are usually new & shiny too.

Now imagine you are part of a team that wants to solve BIG problem, over a LONG timescale. Suddenly the picture becomes more nuanced as software features aren't the only thing that matters anymore. Would you really want to make the success of your project dependent on a third-party software company whose objectives aren't aligned with yours and that can take the tools you use away from you on a whim? FOSS makes a very strong point there. The interface may not make you cry and melt your eyes with shiny buttons, but you can be sure that years down the road the software will still be there and you can read your old files no matter what, as the file formats are open. You can depend on the software, and,.more importantly, you can assess and proactively mitigate the failure modes in advance! (E.g. the tex files from the 90s that you can find on some researchers website can still be compiled today; I imagine the entire research community would be significantly hindered if they would have to deal with the zoo of formats of proprietary software, rtf, doc, docx etc. In this sense FOSS can even act as the catalyst for establishing a standard, as is the case for tex files in domains of science.)

Though I agree that for image editing specifically, typically projects are probably not longer than you few months, so in this you case I imagine you could go without issues with the flow of new and proprietary software.

So how do you manage your banking and tax issues without going insane? Is your company providing you with high-quality tax advisors that help you deal with this issue?

Diversifying your banking lowers your risk or being locked out of an account, but increases the risk of data and identity theft somewhat, as various digital copies of your IDs and other data now reside on even more servers, creating a larger attack surface in case of a breach of one of the banks.

Is the NI really a must in UK? I have heard of people working there that are working only for a few years in academia who don't have an NI - or at least so they claim.

One year (one day actually) in a decade comes close to the very definition of a black swan event. :)

And yes, people will opt for convenience, not rational behavior:

In some cases, that black swan event will cost more than the cost of inconvenience. For example in the US, it is "inconvenient" to retrofit buildings to make them earthquake-resilient, but when the earthquake black-swan hits -and it will hit for sure, the only question is when- damages will be huge, and costs as much as 4 times higher than investments in earthquake-resilience today: https://www.optimumseismic.com/earthquake-preparedness/what-...

I'm sure Kahneman & friends have a name for this cognitive bias that somehow makes it hard for humans to correctly assess the risk and cost for black swan prevention (sometimes, because of the rarity, these computations in principle can't be made). This type of cognitive bias seems also connected with difficulties humans have in thinking on time scales that exceed their own life spans ...

This. Tax laws are very complicated and the 6 months rule is more a rule of thumb than a "hard" rule. In practice, the tax authorities have a set of tests they perform, where the time spend in a country is just one item among many - and these rules vary from country to country.

I happened to be in Austria, when this happened: https://orf-at.translate.goog/v2/stories/2204205/2204206/?_x...

If you had a bank card from a certain big Austrian bank ("Erste Bank"), that day you could not pay by card, nor get money from the ATM; basically you were locked out. Safe to say, chaos ensued for a number of hours, as many people had too little or no cash with themselves. I remember being at a cantine where a long queue had formed with people with the trays wanting to pay and cantine staff running desperately around with "name lists" to register people in return for their promise to pay when service resumed, which it did in the afternoon.

A cashless society is much more prone to black swan events.

Surprisingly I was told by acquaintances that the incident didn't make headlines the next day, and was casually mentioned among other political scandals.

"Maslow's hierarchy of needs considering the wide picture of house prices, home ownership and the wider changes seen in UK society."

What do you mean by that? I cannot immediately see how house prices and home ownership issues drive diminishing banking privacy .

How can you skip (at least parts of) KYC requirements these days?

To me it seems KYC gets ever more pervasive: I had opened a bank account 5 years ago in the EU/UK space and 4 years ago closed it again. Now I opened an account again at the same bank - and the process was significantly more involved, more documentations needed to be provided for the same service, even though I had been their customer before.

The KYC requirement makes me feel uneasy from a privacy point of view: If it would be an eyes-only verification, I would be happy to provide a lot of data to prove I'm not a bad guy. But since the data gets stored and potentially forwarded to third parties, this significantly increases my risk for data and identity theft, as number of increasing data breaches show: https://www.securitymagazine.com/articles/96667-the-top-data...

And there is also a long-term risk associated with excessive KYC data hoarding about individuals: The atrocities in Nazi Germany were in part possible because the government gathered data about the Jewish population (e.g. by enacting essentially KYC-like requirements for its citizens; though I guess through this lense the word should be "KYJ") and then subsequently used that data to round them up: https://encyclopedia.ushmm.org/content/en/article/locating-t...

Also, what do you mean by "not out of bounds for tens of percent of the retail portfolio of many banks"

This comparison is misleading, because it is based on a wrong perspective of how research works:

In programming, if everyone would "just" be good enough, everyone could the elusive 10x programmer. In research, even if everyone would be the equivalent of the 10x programmer, there is simply often a lot of "boring" work to be done, such as what you dismissed as incredibly incremental contribution, to prepare the foundation for the next breakthrough. That means that "10x researchers" could, for various reasons including luck, end up in the 100th percentile because they have chosen a field that is not yet ready for a breakthrough (think, e.g. ML researchers slogging through the AI winter, wo didn't stick with ML until hardware became fast enough). Being the person making the groundbraking achievements means standing "on the shoulder of giants" as someone said a few centuries ago.

I wouldn't be surprised if today there are more "10x researchers" out there than there are breakthroughs waiting to be made.

I work in theoretical machine learning and what I don't like about this piece is how casually the author assumes that what goes us to the present state (hacky approaches where you converge to something "good enough", unless the model is very simple, e.g. a feedforward net---and even here there are still open questions---and you have some theoretical guarantee) will also get us to "AGI".

I believe that is unlikely. Here's metaphor for that, that I often like to use when I speak to engineers doing empirical ML work: In ancient times people build a lot of nice structures, such as pyramids, cathedrals etc. By trial and error many rules of thumb were devised and they more or less worked. But it's safe to say things like earthquake-resistant skyscrapes and modern bridge cannot be build without deep theoretical insights into structural mechanics. These are highly optimized, intricate structures. The same is probably necessary to build highly optimized, intricate models that deliver what we now consider to be "AGI" - but then again, the world of ML is full of surprises.