Vetch

With high probability, what's happened here is this code is an important piece of code-infrastucture in that it's copied into a fair number of places. Which means humans are copying it without attribution or downstream of someone who did while relevant license is not propagated anywhere near as reliably.

It doesn't change licensing issue but it does mean people are already copying and using copyrighted code without respecting original license and no AI involved.

There should be a way to reverse engineer code LLMs to see which core bits of memorized code they build on. Another complex option is a combination of provenance tracking and semantic hashing on all functions in code used for training. Another option (non-technical) is a rethinking of IP.

This scenario is specific to neither github nor copilot. It will always happen for any combination of a code generating LLM trained on all publicly available code.

Unfortunately, Copilot is a lot more capable. Most important is that it works with many more languages out of the box, is continuously updated, has more mathematical plus scientific knowledge and is better at understanding your comments.

As for which to use, you'd want to take the largest model that can you can fine-tune (not necessarily on your machine) and fits comfortably in your machine. The small models aren't as flexible and are basically just a cleverer autocomplete. Copilot is a lot smarter than that and smarter than all models I've tested up to 6B. It can often cleverly work out a lot of local patterns for what you're doing and this is true surprisingly often for novel languages, math, understanding graph notation, categorization policies and much, much more.

I'd gone in hoping I could remove my dependence on Copilot but left disappointed. The small models up to 6B just don't compare. I don't have the resources to run the 13B model which is probably behind anyways due to not having trained as long on as much as copilot (which is continuously updated). I also suspect Copilot has a great deal of hand-coded hacks and caching tricks to improve the user experience.

I think your point is more interesting but the problem is tabula-rasa knowledge starts. A human isn't born knowing about quantum mechanics, christoffel symbols or what pushforward measures are. If there was just a method to learn facts from scratch as cheaply as brilliant humans do, it would be so amazing. Even if you count from elementary school years, humans still end up with less energy spend by several orders of magnitude.

Transformers themselves are a lot more effective compared to n-gram models or non-contextual word vectors. I imagine there is something to Transformers as Transformers are to word2vec.

The key issue is learning effort (such as energy vs time). Congenitally deaf-blind humans with no accompanying mental disabilities as a shared cause can learn as children just fine without any video or sound from comparatively low bandwidth channels like proprioception and touch.

Another issue is what we really care about is scientific reasoning and there, if anything, nature has given an anti-bias, at least at the level of interfacing with facts. People aren't born biased towards learning Metric Tensors and Christoffell Symbols but it takes only a few years at a handful of hours a day using a small number of joules for many humans to get it (I'm counting from all grade school prerequisites vs GPUs watts x time). Much fewer for genius children.

Have to temper expectations with fact that a generated video of a thing is also a recording of a simulation of the thing. For long video, you'd want everything from temporal consistency and emotional affect maintenance to conservation of energy, angular momentum and respecting this or that dynamics.

A bunch of fields would be simultaneously impacted. From computational physics to 3D animation (if you have a 3D renderer and video generator, you can compose both). While it's not completely unfounded to extrapolate that progress will be as fast as with everything prior, consequences would be a lot more profound while complexities are much compounded. I down weight accordingly even though I'd actually prefer to be wrong.

I remember reading that geniuses tend to be more likely to have creative or artistic hobbies. Art is an instance of symbolic abstract reasoning. The level of originating new modes of expressing shape and form, which is what all the earliest humans would have had to do, is the highest mode of intelligence.

You can't use today's standards where it's all streamlined and commodified (though still requiring talent) on that period of time where every little innovation, even simple counting, would have required incredible brilliance and lateral thinking.

My experience with the APIs is Google is excellent and Microsoft is slightly better. And the offline model I've been using that's nearly as good as both is facebook's wav2vec2-large-960h-lv60-self.

Don't believe what's on marketing pages, they rarely transfer to the real world. Will have to make time to try it and see. In theory, given task diversity and sheer number of hours, it should be a lot more robust but will wait on evidence before believing any claims on SoTA.

An automatic rifle has almost no positive use to it. This software is not like that and more like something that makes everyone better at a skill, uniformly raising the artistic talent floor. The large majority benefit but some would use their new earned skill for bad. It would be like denying literacy and the printing press because there will always be some people who use it to write very bad distressing things, manipulate more easily en masse and spread misinformation.

Perhaps it's the use of the AI moniker but this thing is a computer program under the control of a human, who is the person responsible. It takes time and effort to use and is very much not like automated elevator control. In order of difficulty:

1) figuring out how to rejig the prompt to get what you'd like, adjusting seeds and tuning configuration options. The more control you want, the more complex and manual the pipeline of backing software will be. All it does is amplify everyone's innate artistic talent.

2) Coming up with a good prompt. This relies on the person's imagination, facility with words and familiarity of limitations of the image generating software and its training datasets.

3) Selection. This can be a frustrating experience that tries one's patience.

> made a very pornographic image. Who would get the blame?

You would, it's not like the software automatically distributes all its generations. The vast majority of images these software generate are not good enough to be shared and aren't. You made the conscious decision to share it.

Even if it were an AGI, you would be responsible. It's very much possible to commission a ninja from a human artist and get something very pornographic on a famous celebrity and you would be held responsible for choosing to share it, you had the choice not to.

The article is thorough and covers why your [2], under tweening, is insufficient. Video beyond stereotyped repetitive movements is likely AI-complete.

Presumably you'd want to be director and be able to control camera angles. You might want to have a cartoon or 3D render visual style. Unreal engine prompt tag even more literal. Most important, you will want actors. Anything alive will have to move convincingly and intentionally. Some might even bring up the ethics of allowing the movie to end.

Before proper video, I think we'll first have to see tools for music, 3D rendering and animation that bring down difficulty by orders of magnitude.

Being differentiable does not make something a neural network nor does it mean the differentiated function is meaningful (such as can be the case in highly branching or otherwise complex control flow). It also doesn't always escape combinatorial explosion and that fixed structure limited precision strictly bounds how far a neural net could look ahead.

By making something differentiable, you're hoping relaxations or some smooth proxy doesn't break fundamental structure, allowing you to solve an easier problem than discrete combinatorial search. Sometimes there can be barely any leveragable structure. This has proved very difficult to achieve in general and is a sort of holy grail in some fields.

That said, there are powerful game playing AIs that use either no (DeepNash) or minimal (DORA) search. But in general you can't get something for free, you're always paying something with an approximation.

I do not believe this is still true in MuZero, where no rules are ever explicitly encoded, beyond what leaks in from the reward function?

If you look behind logic programming at what powered deduction engines, you will find that its core is very much tree search algorithms like depth-first backtracking search or iterative deepening, which IIUC, is also central in chess engines.

The rest is basically regular programming in that its largely bound by how well the programmer models. For a large number of things, the scope of what a human can consider lags an optimization process. In chess, it at first seemed like this might be the case there too, but latest performance points to maybe not this time.

And it failed at that. As the text gets longer and longer, the lack of synthesis across the contexts you've glued together becomes ever more glaring. It's a decent hack but not a solution. More research is needed for how to forget properly across long contexts.

Maybe if you've read the story it'd seem good but as someone who hasn't, I consider it a poor summary, especially the latter half.

The original phrasing could be worked on but it's not so far out as you say either.

> passing blame for current problems

Humans think we're special but we're still slightly more complicated networked state machines. This means the issue of metastability, where a pathological state is entered and maintains even with the triggering stimulus removed, due to sustaining effects such as fitness criteria given rampant corruption, is real. Control and dynamical systems theory is as applicable to human systems as it is to traffic networks or distributed systems.

This isn't to say current problems are solely because of past inequities but worth recognizing that once entered into, exiting by solutions which entail solving challenging coordination problems are extremely difficult to obtain.

> black-and-white-and-rose tinted glasses

I know next to nothing about Indian history but what they says sounds plausible even if causality for present is more involved. I have at least heard of the famous Indus Valley Civilization and its emphasis on baths, scientific precocity, exceptional levels of egalitarianism, pacifism and influence on subsequent Indian civilization. https://en.wikipedia.org/wiki/Indus_Valley_civilisation#Post...

What is a nudge? Do dark patterns, tactics employed by casinos or loot box games that depend on whales and children of wealthy parents count?

Hi, I'm looking but can't seem to find instructions on how to do tokenization. Where is spm model, is it "flores200_sacrebleu_tokenizer_spm.model" or something else? And is it direct or spm -> dict? Or how to prime model for a specific language pair?

Godot is a great engine that will continue to grow and improve. Unity too is excellent, with many useful professional and tutorial assets. Unity's 3D offerings and tooling are more mature, even as an indie. Some of my favorite professional games made in Unity are: Dyson Sphere Program, Cities Skylines, Pillars of Eternity and Disco Elysium. These are all elite tier for their genres.

Regarding tooling, there's a lot of cruft sure, but that's just Sturgeon's law at play with Unity's size and age. Another aspect of Unity's reputation is it faces a fate similar to web development: it's a lot of people's first experience with programming.

The Godot experience is cleaner, more streamlined and constantly adding more great features. Godot is newer compare to Unreal and Unity so hasn't had time to accumulate cruft. But as long as developers aim for some amount of backwards compatibility, it will happen. That's not necessarily a bad thing.

The interesting aspect of GPT-3+ is their reasoning like behavior such as Chains of thought or "step by step" generation. There's no evidence RETRO as is has sufficient capacity for the actually interesting behaviors.

Github's position doesn't appear to offer any advantage with regards to Copilot's creation.

OpenAI Codex (which copilot grew out of IIRC), Amazon and Salesforce versions of Copilot exist. Huggingface Bloom was trained on a sizeable amount of public code. Tab9, now behind, was one of the earliest to combine public code repositories with Deep learning for smarter autocomplete. The data requirements for Transformer scaling mean any and all public facing repositories will be assimilated, whether Github, Gitlab, Stackoverflow or so on.

Wish more energy was spent on how to fund pretrained models that will also run efficiently on CPUs, fine-tuneable to one's language and local environment. Removing reliance on cloud services.

Curious about people's opinions on Dall-E 2 or Google Image-gen, which parallel pretty much the same thing with Renders, Illustrations and Paintings, or upcoming models doing the same for voice acting and music. Coders seem more excited about the potential of those tools.

There is an economic incentive to detect machine generated output and curate trusted sites since the feedback loop of training models on an unfiltered internet of mostly generated junk output will eventually lead the models into converging on a useless degenerate state.

The vast majority of generators of fake news today are from Content Mills, Copy Paste Mills and general SEO spammers. Political misinformation is the other big generator. The economics of it and not "ethical" gate keeping is what will affect their output. Realistically, normal people don't have the ability to coordinate hordes of proxied IPs to attack social networks requiring account sign ups and custom feeds.

The value of exercising critical thinking, checking trusted curated sources, information hygiene, recognizing and avoiding manipulation tactics in news and ads will have to go up. The internet is already almost entirely unreliable without requiring any fancy AI. The listed skills will be necessary regardless any increase of politically manipulative content, advertisements or product misrepresentations.

The experiment is okay, it's actually a special case of a concept explored in Egan's Permutation City and your observation about what's really doing the work applies there too. Except story goes in unsettling directions by really taking noise generator aspect seriously. Things get interesting when sections of patterns become self interpreting.

A similar thing could be done for brains: record with necessary accuracy, all voltages, membrane potentials and any key biochemical concentrations. This will take a finite number of bits. Look for a decoding of recorded data from heated iron bar, convert those readings, instead of using original, and play that back into state clamped brain. Does being able to read conscious state into brains from hot iron invalidate them too?

Another relevant story is Wang's Carpets. We might look at some alien moss or fungal mat and think it primitive. But later our technologies and knowledge advance to the point we can now see it's running a complex computation with intelligent agents inhabiting. Did the creatures not exist until we could decode them?

One of its pivotal flaws is:

> Since there is no definition of computation without reference to an external observer, a system in isolation just cannot compute, which suggests that a conscious being cannot compute.

This is an assumption they do not try to and cannot prove. It's also what much of their argument rests upon.

Related ideas are subjectivity of emergence or what counts as an observer for Wigner's Friend.

Trust is an ideal to strive towards but the unfortunate case is that its reliability quickly starts to strain with scale. I'd argue the economy at large is held together not by trust but by reputation, curation, regulation and threat of lawsuits and jail time. With the last concentrating trust in legal institutions, overall reducing trust burden.

In many countries, the richest companies operate unethically, with impunity because the legal system is in the pocket of a corrupt few in power. In such countries trust in institutions is nonexistent. In developed economies, regulatory capture still leads to inequities and gross inefficiencies.

Today on the internet, reviews and product search results can no longer be trusted. For a very long while, buying from a large class of products on Amazon was a gamble. A breakdown in reputation and curation.

Computing environments were originally high trust, a property they could not retain at scale. Encryption, https everywhere, anti-ad tracking, secure VMs, gated app stores, "Don't trust the client" designs, onion routers, anti binary blobs are technologies and principles which reduce the need for trust. Technologies like DRM, copy protection rootkits and Trusted Computing are corporations communicating their lack of trust in humans inhumanly.

There's always some minimal amount of trust necessary. The optimal minimum generally decreases with scale. Hawala might seem an exception but closer inspection shows it depends on shared belief systems, local interactions on reputation networks and honor; memes which reduce trust burden at scale.

I'd also argue bitcoin isn't about trust but about issues which arise with open permissions and no centralized enforcement mechanism. Its throughput issues addressed by "layer 2" solutions, as decentralized credit networks (from a dispute and authority perspective), actually have fair overlap with honor based informal money transfer systems but without the centuries to have ironed out issues.

Being an agent isn't about how complex the model is, nor about how much RL is used. Bacteria is an agent. Sponges and Jellyfish are agents. It's about if the model engages in exploration and or generates inferences in service of non-trivial control problems.

> On the other hand, if you strap GPT-3 on an agent in a 3D Home Environment

GPT-3, just as any probability distribution, can inform an agent's actions but that doesn't make it one. WebGPT3 is an agent though.

The other ones are smaller but not much worse according to their tests (oddly, in the Winograd Schema Challenge and Commitment Bank tasks, the largest model actually appears to be worse than much smaller ones).

30B parameter models are already large enough to exhibit some of the more interesting emergent phenomena of LLMs. Quantized to 8 bits, it might be possible to squeeze into 2, better three 3090s. But the models also seem undercooked, slightly to strongly under-performing GPT-3 in a lot of tasks. To further train the same model is now looking at > 100 GB, possibly 200GB of VRAM. Point being, this is no small thing they're offering and certainly preferable to being put on a waiting list for a paid API. The 6.7B and 13B parameter models seem the best bang for your buck as an individual.

Such failures are interesting, in the physicist's sense.

That it struggles with composition sometimes but not other times might indicate model has learned a particular distribution factorization which is good much of the time but makes it unable to properly account for conditional dependencies in certain queries. The issue of losing track of lots of details may also be related.

Its struggles with negation is also fascinating because that has been a thorn since the early days of GOFAI. Negation is trickier and more involved than what one might naively assume.

For those taking relief from such weaknesses, it's worth remembering that this is as bad as it is going to be and probably not for long. Model size is surprisingly small when compared to recent behemoths. That it can already do all that it does is certainly cause for concern.

I would not say that. Facebook, Microsoft and Google release plenty of useful models. EleutherAI have released 6 billion and 20 billion parameter language models. Huggingface has been training a 176B model [1].

The issue isn't a lack of models or data, it's that larger models are impossible to train without paying hundreds of thousands to millions of dollars. The hardware requirements for simply running the models already prices it out of reach for most.

These models are rather powerful but the immediate future is one of accessing them by cloud services. GeForce GTX 1080 Ti was 5 years ago, since then memory has roughly doubled in consumer GPUs. To run the highest end models on single GPUs, HW will need to 20x to 70x in memory at the same time as serious gains in flops/Joule.

I suppose improvements in CPU parallelism and RAM speeds will also go a long way towards making such models runnable on reasonable consumer hardware, albeit at slower speeds.

[1] https://huggingface.co/bigscience/tr11-176B-ml-logs