Unbox Research

We write articles about AI! We’re a machine learning research & development company. We find answers in data. Founded by @tylerneylon.

[Unbox Research @unboxresearch]

I used to laugh at Hollywood whenever an actor said "enhance image!" and a blurry pixelated mess morphed into a crystal-clear portrait of a villain (a process called upscaling). Now, at long last, it's Hollywood's turn to laugh at me.

[article: https://t.co/4DbDUr5vRk]

Surprisingly, it's not because of my humorous-yet-informative linkedin posts — but rather because a few weeks ago a team of researchers introduced a new upscaling algorithm they call SUPIR (Scaling-UP Image Restoration) which can actually enhance images fairly well.

SUPIR uses a plethora of fun ideas that, together, balance out to a great combination of semantic awareness and fidelity to your input image. They start with the SDXL model to perform image generation, but with heavy modifications so that it will act as an upscaling system rather than a text-to-image generator.

One trick they use is to invoke another model to derive a text description of your low-quality input image. This text description can then be fed into SDXL's text input to help condition the generation. This isn't enough by itself to produce a faithful higher-resolution image, but it helps by explicitly trying to understand the semantic content of the input photo.

Another technique they use is to integrate SDXL generation with a modification of another network called ControlNet. ControlNet is a system specifically designed to be integrated with Stable Diffusion (SDXL is a variant of Stable Diffusion) so that image generation can be further guided by additional external inputs.

The authors also provided a few additional tricks. For example, they had to figure out how to map a low-quality image into the same latent vector as its high-quality version since SDXL "thinks" in terms of latent vectors. They also had to figure out how to avoid catastrophically damaging the output from SDXL when training the ControlNet additions. See sample results, and read the full summary on Learn & Burn.

Work by Fanghua Yu @JasonGUTU Zheyuan Li, Jinfan Hu, Xiangtao Kong @xinntao et alia.

[Unbox Research @unboxresearch]

I've written a few times before about learning when text is written by an AI.

[article: https://t.co/vLWeeocrFV]

Up until now, the best technique I know of has been to _edit the LLM_ to secretly give itself away — referred to as adding a watermark. But why would a nefarious ne'er-do-well use a watermarked LLM? They wouldn't!!!! (Cue scary violin music.)

Fear not, gentle reader — there is a glimmer of hope in this dark, cruel world, where plagiarists traipse about unchecked and misinformation infiltrates our news feeds with nary a second of skeptical scrutiny. A new paper, posted on arXiv in January, has been able to achieve an astounding 97.6% positive recall rate (percentage of fake papers it identified correctly) with a scant 0.01% false positives on student essays. As one point of comparison, a competing model called GPTZero was only able to achieve 64.7% recall in this same setting.

The coolest thing about this LLM-detection model, called Binoculars, is that it does _not_ have to be trained to detect any particular brand of LLM — it distinguishes between human and _any_ AI. At a very high level, its secret to success is a new formula to compare an LLM-vs-LLM distance against an LLM-vs-input distance. If the LLM-vs-LLM distance is much smaller, then it declares the input to be human-written; intuitively, things close to LLM-based text are seen as also LLM-based, and farther text is seen as human-written.

I've skipped over the fun parts in that description. It's the technical details that make this paper cool. The authors use an existing measure of LLM quality called perplexity, and then extend it to invent a new measurement called cross-perplexity (analogous to cross-entropy). Read Learn & Burn to learn about this exhilarating new quantity, and why they call their new technique _Binoculars_.

Work by @ahans30 @A_v_i__S @ani0075saha @micahgoldblum @jonasgeiping @tomgoldsteincs et alia.

[Unbox Research @unboxresearch]

The movie Tron was released in 1982, and made the first "extensive" use of computer generated imagery.

[article: https://t.co/9OppozrkkR]

"Extensive" in this case means only 15 minutes of full CGI. The blending of live action video with rendered images was an early encounter with the uncanny valley of simulated reality — a valley that AI is currently dominating with its weirdly wrong text-to-video results.

Google Research is working on this weird wrongness — to sound more academic, they're trying to improve the _temporal consistency_ of text-to-video output. A couple weeks ago, they released a new model called Lumiere that integrates some natural and effective ideas in this direction. One of the things I like about this paper is that you walk away thinking "why didn't they do this in the first place?" (I love it when papers open your eyes like that.)

Previous text-to-video models had a habit of generating some time-separated low-resolution key frames, and then using side algorithms to upsample and to fill in gaps between them. Because these were essentially separate algorithms that had different levels of understanding about the context of a scene, temporal consistency wasn't great (the videos looked jumpy or unreal). Lumiere improves on this by directly incorporating time as a dimension into a generative U-net. You could think of previous models as generating a few disparate 2d key frames of a video, while Lumiere is good at generating an entire _continuous_ 3d object in one go.

The Learn & Burn story walks you through this, from Will Smith eating spaghetti, through a cool idea called X-T slices, over to a stronger explanation of the new STUNet (space-time U-net) architecture.

Work by @omerbartal @hila_chefer @Roni_Paiss @ShiranZada @arielephrat @JunhwaHur @MikiRubinstein @oliver_wang2 @DeqingSun @talidekel @inbar_mosseri et alia.

[Unbox Research @unboxresearch]

If you were to put an open source LLM into production today, which one would you chose?

[article: https://t.co/D5cKmJ5ojW]

Many people jump straight to llama-2, perhaps opting for the large 70b version, which has the best performance of the different sizes available. However, this model needs more expensive GPUs, and will take longer to run on every input because of its large size. It feels like we're faced with a tricky choice of either good performance but high cost and low speed, versus worse performance at reduced cost and higher speed — achievable by using a smaller model.

This is all leading up to a model just released in December called mixtral 8x7b. It kind of sounds like the model has 56b parameters, but that would be hiding some key facts. The reality is that this model is a mixture of experts — meaning that it internally has multiple submodels, and each time you run inference, a different subset of weights is used. The result of this technique is that (since only two experts are used at a time), you only need enough memory for a 13b model. (If it bothers you that 13 ≠ 2×7, that bothers me too, but I believe the "7" is a rounded up "6.something.")

Here's the fun part — because the submodels used are chosen carefully (and chosen based on traditional ML training), you still have performance much closer to a large model. In fact (and this is where things get crazy), mixtral 8x7b _outperforms_ llama-2-70b on some key benchmarks. In other words, we have an open source model that uses less memory (so is faster and cheaper) than llama-2, and actually gives us even better quality results.

You don't have to take my word for it. Search for either of those models on huggingface's LLM leaderboard: I see the top mixtral model outperforming the top llama-2 model there. Want to know more? Here's the full write-up on mixtral 8x7b, possibly your current best-choice for an open source LLM.

Work by @AlbertQJiang @alexsablay @arthurmensch @Bam4d @dchaplot et alia.

[Unbox Research @unboxresearch]

What's the next challenge for AI image generation? Well, what if we asked a model to generate a single image that simultaneously includes two separate scenes when looked at in different ways?

[article: https://t.co/WS7cTt33aN]

It's basically a magic trick. One way to do this is to draw an image that looks like one thing — say, a typical high school student from Brooklyn — when you view the image right-side up, but when you turn it upside down, you see something entirely different in the same image — say, a friendly super-powered neighborhood crime-fighter.

The authors of a recent paper showed how to modify Stable Diffusion to produce exactly this kind of optical magic: You provide the kind of transformation you want to use (such as "flip upside down"), and text descriptions of what each image perspective should depict. The algorithm does its best to generate that image — and, in general, I think it does a great job (for a version 1).

One of the creative aspects the authors tackled was to exhibit a wide variety of transformations — beyond just "flip the image upside down." One of my favorite is a puzzle style, in which the algorithm generates traditional puzzle pieces — but there are two different ways to put them together, and the two different ways to solve the puzzle reveal two separate images. Obviously, the colors used in the images need to be the same, but the objects depicted can be independent.

The authors also worked with a color inversion transformation. The simple black-and-white version of this is to flip white pixels for black pixels, and in general to replace the lightness value x by 1-x (so dark gray becomes light gray, for example). One of my favorite images they generated uses this transformation, and I've included it here. It might be hard to recognize at first that it's actually the same image (the grainy man/woman photo).

How did the authors do this? Basically, they iteratively kept moving each perspective toward the goal scene for that perspective and then combining the small steps back into a single image. Find out how they modified stable diffusion to pull this off in this week's Learn & Burn.

Work by @dangengdg @invernopark @andrewhowens

[Unbox Research @unboxresearch]

What stands between us and widespread adoption of LLMs in our day-to-day lives?

[article: https://t.co/gPuJ3BRpZ9]

In one sense, we're already seeing hints of AI applications: email portals offer to write emails for us (no thanks), and I've tried out tools that summarize work calls for me (yes thanks). At the same time, my smart speaker seems awfully dumb since I started using ChatGPT.

I think the main point of friction (ignoring the need for improvements in accuracy) is the _cost_ of running an LLM, which is exorbitant. For example, if you want to serve an LLM that uses 80GB of GPU memory, then you might buy a fleet of H100 processors, with an up-front cost of $30,000 per H100 — and we haven't even gotten to the recurring price of powering, cooling, and using this hardware in production.

This is all leading up to: Both edge computing (running an LLM on your phone or speaker), as well as making use of cheaper GPUs are attractive. Perhaps surprisingly, the primary differentiator between an expensive GPU and a cheap one is _memory_. Historically, if your model couldn't fit into your GPU's memory, you simply couldn't run it. But that just changed.

Researchers from Apple (!) have actually published a handful of techniques that, taken together, make it practical to run large models even when they don't fit in memory. How is that possible? Essentially, they've figured out how to be clever about which subset of parameters they bring in to RAM. With each iteration of loading/unloading certain weights, they only transfer a small amount of data, and the loaded data is sufficient to provide a solid approximation of what the full-data model would have provided.

How did they do it?!? The technical details are covered by the Learn & Burn team this week.

Work by @KeivanAlizadeh2 Iman Mirzadeh, Dmitry Belenko, S. Karen Khatamifard, Minsik Cho, Carlo C Del Mundo, @morastegari and @MFarajtabar

[Unbox Research @unboxresearch]

Within the research world, Gaussian splats have proven to be an unexpected breakthrough to empower realistic, data-driven 3d models.

[article: https://t.co/Q4g5WV2T43]

I can't wait to see these techniques integrated into real systems. My guess is that Meta will be a pioneer here — in particular, that within the next couple of years, we'll see realistic avatars available through their VR line.

This isn't such a wild guess. Meta has been releasing a series of increasingly impressive avatar demos and papers. They provided a VR-based interview with Mark Zuckerberg in September 2023 — calling it the "first interview in the metaverse." Learn & Burn has covered some key research within Meta supporting these avatars. A previously covered paper provided new ways to accurately model moving bodies with independently-modeled clothes (to render things like realistic wrinkles, and non-frozen cloth material).

More recently, toward the end of 2023, a newer paper covered highly realistic face rendering. One thing I want to emphasize about this research is that these avatars are "data-driven" in the sense that they're typically based on 2d images and 2d video. So, once we have the AI-based (optimization-alg-based) methods to go from 2d-to-3d, then we can enable consumers to build their own models, such as their own avatars. This is part of the breakthrough.

The newer paper is motivated by the fact that we notice a *lot* about human faces. It's the most important part of the avatar. So the standard Gaussian splat model is modified by a special component that knows how to encode and render a facial expression that's been vectorized from images of a real person's expression. The modifications support not only a highly nuanced range of facial expressions, but also the ability to relight the scene at will — an ability that the original splatting method was not designed for.

Read the full coverage of the paper in Learn & Burn.

Work by @psyth91 Gabriel Schwartz Tomas Simon @JunxuanL and @giljoo_nam

[Unbox Research @unboxresearch]

One difference between a human brain and an AI model is that organic neurons don't need to load data from RAM.

[article: https://t.co/AKdsaxdQAt]

Even when the weights of an AI model are loaded into GPU memory, there's still some load time to get data from that memory into the core that's about to run some calculations. This makes me wonder: What will be the next big breakthrough in model efficiency?

Inspired by that thought, I want to tell you about a pair of great papers from the end of 2023 written by two authors (Peter Belcak and Roger Wattenhofer) @ETH_en. They've invented a radically more efficient replacement for feedforward networks, aka vanilla matrix-multiplication layers of a deep neural network; they've also tested out the theoretical power of their idea in an LLM.

Here's their idea: What if we replaced the work of a feedforward network by a binary decision tree? By _binary decision tree_, I mean a tree that receives a vector, then has a set of nodes that output "left/right" based on the original input vector, and end with leaf nodes that can take that original input vector and output a modified vector. Such a tree can be faster when (a) each node is fast, and (b) each leaf is efficient, providing a smaller output vector than the full-sized feedforward network it replaces. Keep in mind that a decision tree with about N nodes only needs to compute about log(N) internal nodes to provide its output.

A good question at this point is: Even if this is more efficient, why would we intuitively expect to arrive at good quality results by using fewer learned weights? After all, using fewer weights is like using less of the information we gathered during training. But, in practice, it's usually only a small fraction of neurons that are actually providing value for any given input.

If some neurons were never providing value (eg, always outputing values near zero), then we could just drop those neurons. But the reality is more complex. Often, any given neuron will provide a near-zero value on some inputs, but then fire strongly (have a strongly non-zero value) on other inputs. So we can't throw out the neuron, though we know it can be ignored in many cases. (One piece of evidence this is true: Consider how well 4-bit quantization works. This means that it's more useful to say yes/no that a neuron is active for a given input than to know its precise value; and most of the activity information is in this yes/no rather than in precise values.)

If we can often ignore a large fraction of neurons for any given input (but which ones we ignore depend on the input), this is a good fit for decision trees; they're exactly the structure that carefully categorizes an input, and then makes a final call based on that categorization.

Find out how these trees are architected and trained in this week's Learn & Burn.

[Unbox Research @unboxresearch]

I want to kick off the new year with thoughts on 3d avatars: My favorite aspect of VR is a sense of presence when spending time with other people.

[article: https://t.co/5iKwk4K9Oe]

The technology so far doesn't offer a lot of this. If you have a Quest headset, then you can play minigolf with other people in Walkabout, or compete in a game like Beat Saber.

The current avatars in such games are extremely simple: People don't have legs (the headset doesn't know where your legs are), and faces tend to be simple and abstract, without changing emotions. Even with these drastic limitations, it's fun to play with other people because you feel much more like you're hanging out with actual humans than you would if playing 2d games. (Ignoring voice chat for now, which is a channel independent of 2d-vs-3d.)

We'll soon be entering an uncanny valley for 3d avatars as they begin to look more like complete humans. At the same time, a wave of impressive new technology is opening doors to brand new real-time techniques. I bet that a few years from now, we'll see much more realistic and emotive avatars in VR-based worlds.

Why is this technologically difficult? From one perspective, we're trying to do what Hollywood has done with highly involved mo-cap techniques and make it available to consumers equipped with hardware light enough to be held up by their noses. I'm still not sure how VR companies can solve the full problem of rendering emotive, dynamic, and personalized avatars, but research from a new paper is moving forward on one part of the puzzle: We can now build models that look realistic and can be rendered quickly from a video of a person.

One core idea behind the paper's technique is a way to make a Gaussian splat model dynamic — allowing it to move as a person moves. This is a fairly new idea, since the already-new Gaussian splat technique was originally built for static scenes and objects. The paper goes into more detail than just pose matching, and shows careful consideration of rendering clothing that moves independently from the person's body.

Find out how Gaussian splats can model sophisticated dynamic objects (like people) in this week's Learn & Burn.

Work by @w_zielonka @psycharo @psyth91 @MZollhoefer @JustusThies and Javier Romero.

[Unbox Research @unboxresearch]

What is intelligence? This question is surprisingly practical today, as researchers focus on AI's current holy grail of AGI — artificial general intelligence.

[article: https://t.co/wB5E0b5b2j]

AGI is a central focus of research by OpenAI and others. But it's going to be tricky to say whether we've accomplished it. Suppose you can choose to use either OpenAI's version, or Google's. Both services give examples of problems they can solve better than the other. How do you decide which model is more intelligent? How can you precisely express that level of intelligence to your coworkers, your users, or your boss? Within research groups, how do they know if they're making meaningful progress?

There are a few reasons it's difficult to quantify intelligence. One is that a mere calculator is already superhuman at arithmetic, although it doesn't seem intelligent — so saying "better than humans" isn't right. You might be tempted to say "can help with any problem." And yet ChatGPT is quite general — useful in many different ways — without being better at solving open-ended problems than people are. So neither "better than people" nor "helps with many types of problems" are sufficient on their own.

To begin making their intelligence metric, DeepMind researchers defined a numeric scale of skill-level. They define six levels, from zero intelligence (to include, say, calculators) up to the top level, which is defined as a system better than all humans at a given task. Now consider a tricky case: AlphaZero, a model which seems intelligent in that it can learn to play sophisticated games, but is narrow-minded in that it can't speak or see or do anything besides play games. What level is AlphaZero? In order to express the range of problems a model solves, the DeepMind paper suggests we classify agents as either "narrow" (having few applications) or "general" (having many). Thus, we can meaningfully label AlphaZero as highly intelligent (since it can beat any human), but narrow in focus.

Beyond quantifying intelligence, the paper does a nice job of considering other popular definitions of intelligence. The authors derive from other papers a set of six values they think any good definition of intelligence should cover. For example, a good definition of intelligence should focus on capabilities instead of processes, and the usefulness of the problems being solved.

The full intelligence quantification, as well as the other values enumerated by the researchers, are covered in this week's Learn & Burn.

Work by @merrierm @jaschasd @nfiedel @triswarkentin @AllanDafoe @AleksandraFaust @clmt and @ShaneLegg

[Unbox Research @unboxresearch]

I'm reading an old sci-fi classic called The Moon is a Harsh Mistress by Robert Heinlein. It contains a passing reference to the need for an AI in the story to read virtually every book ever written in order to understand humanity.

[article: https://t.co/uS5FuKXKVN]

Another character responds, saying the AI can read fast enough to do just that. It got me wondering — what can researchers do _after_ a model has read every book ever written?

The funny thing is, we're basically at that point already. Based on some quick googling, there are about 32 million books in the Library of Congress, and books tend to average around 90k words each. That gives us 2.8 trillion words, or about 3.7t tokens. Rumors I find credible suggest that GPT-4 was trained on more than 13t tokens. In other words, we've already surpassed 3x the amount of verbiage in the Library of Congress.

So how can models get better once you used up all the training data you could get your hands on? A new image-gen paper from Google Research uses a simple and replicable idea: Take a human/AI hybrid approach to improving the quality of the training data. This idea may sound simple, but I suspect it may become a common tool in model improvement down the road.

You usually can't use AI alone to improve the quality of your training data (at least not for a state-of-the-art system) because doing that would require a model at least as good as your desired goal. You can use humans to make new training data, but they work slowly (c'mon, humans!). A nice hybrid approach is to factor the problem you're solving into a subset of information humans can provide, and another aspect that AI can amplify without losing quality.

I'll give an example to solidify that concept. Researchers at Google wanted to improve the faithfulness of a text-to-image model — they wanted the images to more exactly match all the details of prompts. They found that a lot of the training data had only low-quality captioning, but they had too many images to label by hand. So they sat down and carefully labeled a subset of images manually, giving these images long, detailed descriptions. That data became training data for a captioning model that could then amplify the better image descriptions across their entire image set.

The results are stark. The new model is clearly superior, both qualitatively (check out the link to see samples) as well as quantitatively, based on how frequently the generated images exactly matched the prompt's requests. Check out the full paper summary and see output images from the new model in this week's Learn & Burn.

Work by Eyal Seglais, Dani Valevski, Danny Lumen, @ymatias and Yaniv Leviathan

[Unbox Research @unboxresearch]

I can imagine a future where advanced mathematics has completely changed. What makes math challenging today is the ability to learn abstract technical concepts, as well as the ability to construct arguments that solve precise logical problems.

[article: https://t.co/IdAvGoSlW2]

I believe an AI will one day possess both of these abilities. New research has made surprisingly good progress in this direction.

Think of the future of art or writing. One negative perspective is that human creators will become obsolete, but I disagree because I still feel amazed by the great works other humans have already made, and continue to create. Another negative perspective is that AI will always be derivative in its creation — for example, images made by AI are still (in my opinion) easy to spot versus human originality. I suspect that, given more research, the novelty of AI-created ideas will improve (as it's been doing over the past year).

Math is similar to art and writing: People have goals they want to achieve, a craft they follow toward those goals, and a sense of taste that appreciates beauty in something well-executed. So, given the concept of an AI that can learn and solve advanced mathematical questions — where does that leave mathematicians of the future?

My guess is that, just as microscopes created entirely new fields of study, AI as a tool will shine light on new areas of mathematical thought. When previous challenges become easy, people don't look away, they look forward: What can we now do that we couldn't even imagine before? My bet is that mathematicians will find new limits for their AI collaborators, look for new kinds of beauty, and be curious about new ways to think.

These thoughts were inspired by a powerful math model called LLEMMA. It's the new state-of-the-art theorem prover among open source math models. And — to be clear — I'm not talking about adding numbers here, but about formally proving arguments to solve higher-level math questions. The summary is in this week's Learn & Burn.

Work by @zhangir_azerbay @haileysch__ @keirp1 @dsantosmarco @McaleerStephen @AlbertQJiang @jiadeng @BlancheMinerva @wellecks

[Unbox Research @unboxresearch]

The first computer imagery on film looks humorously simplistic today. Similarly, the earliest models that generate 3d objects from text already exist, but are low quality.

[article: https://t.co/DcCYnj8CR4]

Looking ahead, I see an exciting future where the creation of 3d models is completely transformed by the future of techniques that are already being used — things like neural radiance models or Gaussian splatting, iterative geometry improvement, iterative texture improvement, all pieces within a symphony of text- and image-aware pre-trained models.

A recent leap forward in this direction: a new text-to-3d model called DreamCraft3D. The most obvious and exciting feature of this system is that it creates 3d models that are vastly superior to previous models. You can ask for a corgi eating a baguette, and you'll get a realistic 3d object of exactly that.

While this sytem does internally use a temporary NeRF (neural radiance field) to begin making the 3d object, it refines it over time as a traditional mesh of triangles and points. My favorite part about this system is that it teaches us how to deftly compose many pre-existing, simpler models and techniques. They're synthesized in a way that mimics a natural, human-like process: Begin with a rough draft and improve that over time. In a sense, DreamCraft3D starts as a naive 3d artist and learns throughout its creation process how to make the object you've requested.

You can see how both diffusion models and NeRF-style rendering work together in the full paper summary in this week's Learn & Burn (out early for Thanksgiving, gobble gobble).

Work by @JingxiangSun42 @zhangboknight @RZ_Shao @YebinLiu

[Unbox Research @unboxresearch]

Earlier this year, researchers showed that gzip can be used as a kind of machine learning predictor.

[article: https://t.co/zdRRZTSsm6]

A few weeks ago, Google DeepMind added a new, very elegant, perspective — they showed a simple and effective equivalence between any kind of sequence predictor (such as an LLM) and compression. Not only did they show the mathematical equivalence, but they also showed that modern LLMs are amazingly good compressors.

Here's an example of how good LLMs are as predictors: Chinchilla 70b (an LLM) is able to remove over 91% of data to losslessly compress 1 GB of Wikipedia articles. As a point of comparison, standard gzip is only able to get rid of 51% of that same data. Yes, LLMs made the data more than 5x smaller than gzip. Even more surprising is that LLMs are also good at compressing binary image and audio data!

One key ingredient to this work is an old compression technique called arithmetic coding. It's complicated enough that I can't explain it simply in a post, but what it accomplishes is taking a known probability distribution over symbol sequences (for example, knowing which letters are most likely to come after any given letter), and using that knowledge to provide a lossless compressed version of any given sequence. One cool aspect of arithmetic coding is that, given perfect information about the probability distribution of the data you're compressing, it can achieve the smallest-possible output when averaged over many incoming messages to compress.

Now I'll connect that with the new DeepMind work. LLMs work by receiving a prompt — a sequence of words happening before "now" — and then providing a probability distribution on the next token. Mathematically, this is _the same_ as a conditional probability distribution — exactly what arithmetic coding needs to use to perform compression. So any LLM, combined with arithmetic coding, is a compressor. The authors go beyond this and show how you can also turn any compressor back into an LLM (although it's not very practical).

All the fun details are in this week's Learn & Burn.

Work by @gregdeletang @anianruoss @duquenne_pa @jordigraumo @liwenliang @LaurentOrseau @mhutter42

[Unbox Research @unboxresearch]

LLMs require an insane level of compute, making them slow and expensive. It would be amazing if top-quality LLMs were fast and cheap to run, which is exactly the goal of new work from MIT/Meta/CMU.

[article: https://t.co/A19ZjJCART]

Guangxuan Xiao and his co-authors created a new way for LLMs to work with text that's longer than its context length. This can often be a small window, such as the default 4,096 tokens for GPT-3.5-turbo, which is something like 6 pages of a book. If you want to work with more text than that, your current choice is to re-start the LLM from scratch, feeding it a new sliding window each time. Using a sliding window works, but is inefficient because it re-processes the same words multiple times.

You could try to speed things up by simply re-using the deep activations of an LLM corresponding to already-seen words that are sliding within a window, but the quality of the output text quickly degrades when you do that. Those re-used internal vectors assume the existence of previous token vectors which have disappeared — in short, this approach doesn't really work. And that's what Xiao et alia have fixed with a new internal tool called an attention sink, which is a quick way to hold onto information from earlier in the large input text.

One of my favorite things about this paper is that it helped me to better understand the efficiency trade-offs of LLM architecture, beyond even the improvements they added. The fun details are here in the Learn & Burn paper summary.

Work by @Guangxuan_Xiao @tydsh @BeidiChen @songhan_mit @ml_perception

[Unbox Research @unboxresearch]

It would be fun to make images of yourself in whatever setting you can imagine: scuba diving, giving a TED talk, exploring a jungle ...

[article: https://t.co/Z5mOYRASjN]

A while ago, I tried the Lensa app, which supports this kind of thing, but it required ten photos of me to start working, and the generated images didn't always look like me. New research has created a novel system that works much better.

What the researchers did, essentially, is to modify the original Stable Diffusion model so that it can learn what someone looks like from a single input image, and then put that person into the image you want to create (you specify the image with a text prompt). Previous models _can_ put new, learned-from-immediate-input faces into image. What's new here is that this model works *extremely well* at generating highly recognizable faces.

To simplify, they succeeded by fine-tuning specific sub-components of the Stable Diffusion model with a loss function that takes into account how recognizable the face is in the created image versus the provided image. I think of this as Stable Diffusion (which already knew how to paint) learning how to really grok faces (which it wasn't good at originally).

The fact that this is a new learned skill reminds me of Oliver Sack's story "The Man Who Mistook His Wife for a Hat," about someone with face blindness (prosopagnosia). The patient in that short story was intelligent, had perfectly good vision, and was able to recognize many objects — but not human faces. The existence of this disorder suggests that humans, like Stable Diffusion, can have a specifically-tuned part of our brain just for recognizing people. Which makes sense evolutionarily!

To fully appreciate this work, you should really check out the generated example images — see the full Learn & Burn summary.

[Unbox Research @unboxresearch]

I sometimes get a sense that AIs aren't perfect, but it's tricky to put my finger on exactly what they're bad at.

[article: https://t.co/bnBs5UHug4]

Well, ok, if a self-driving car can't drive around an orange cone, that's an obvious fail. But I want to highlight some research that systematically finds more subtle weaknesses in game-playing AIs.

In a new paper, the authors took a Go-playing model called KataGo that is considered "superhuman" because it beats all humans. (I guess if it could only beat men it would be merely "superman.") What they did next is to train a new Go-playing AI that includes a copy of KataGo, so it always knows which move KataGo will make. Yes, this is basically cheating, but remember that the point is _not_ to win fairly, but rather to find weaknesses.

Their strategy worked. Their new model, which I've decided to call "KataStop," found a particular playing pattern that is easy for humans to defeat, but which stymies KataGo. Stymies!

My favorite part about this work, besides the smug sense of transitive superiority I now have over KataGo, is that we're starting to learn nuanced error modes for advanced, previously opaque systems. This approach could be part of the journey from 90% solutions to 99% solutions, which is the trickiest part (as any decent progress bar can tell you).

The full details are in the latest Learn & Burn.

Work by @5kovt @ARGleave @KellinPelrine @norabelrose @MichaelD1729 @yawen_duan @svlevine

[Unbox Research @unboxresearch]

I've always loved world-building: starting from nothing and creating a new universe.

[article: https://t.co/ARBU7N2AwG]

Back in 2009, Minecraft changed the industry by procedurally generating a new planet with every new game. There's something fun about exploring a world that literally no one — not even a human designer — has ever seen before. You're the first person to see that cosmos.

Approaching 15 years later, new forms of world-creation and rendering have become possible. In particular, techniques such as GANs (one approach to generative AI) and NeRFs (one approach to 3d rendering with neural networks) open doors to more sophisticated possibilities.

As an example, what if we wanted to generate a city that's much more realistic than Minecraft? Instead of large blocks and pixelated art, can we get more arbitrary polytopes and fine-grained textures? Yes; this is what a new system called CityDreamer does, published in a paper last month.

I suspect that ideas like this — new generative techniques, NeRFs, and the new Gaussian splatting approach can combine in powerful ways that allow us to create 3d scenes without punishingly intricate manual design work. I think humans are still better at this task than algorithms. But I like the vision of tools that remove a startling amount of accidental complexity, bringing creative minds closer to their creations more quickly.

Work by @zjhzxhz @Frozen_Burning @hongfz16 @liuziwei7

[Unbox Research @unboxresearch]

There's a new way to generate speech in anyone's voice, in many languages, with zero training data.

[article: https://t.co/dpbwZJfipJ]

Ok, hold on — let me back up a minute to put some context on this.

Models like GPT are at the forefront of a capability revolution. It's a tool with so much potential, we don't yet know how we'll use it. Yet one pattern _is_ emerging: The generative ability to start with a tiny seed of partial information, and extrapolate to a full-blown realization. We've seen that happening for text and images. While less developed, it's on the way for music and video.

Let's get back to speech. Researchers at Meta have put together a new approach called Voicebox that not only outperforms competing algorithms, but also does so under increasingly impressive constraints. To begin with, it doesn't need _any_ training data in your own (or whoever's) voice. Just as GPT is pre-trained with others' writings, Voicebox has learned from around 60,000 hours of spoken English (and many hours of other languages). It can use a tiny audio sample of any voice, and immediately produce new audio in the same person's voice (basically text-to-speech). I consider the term "zero-shot" a little fuzzy (what counts as "no labeled data?") but this is a good contender for that label.

There's more. Just as some text models can edit or infill your writing, Voicebox can do things like remove a dog barking in the background, insert new sentences to audio, or change a few words.

And, yeah, in case you're thinking this tech could be abused to fake someone's words — that thought also occurred to me! Yet there are great legitimate applications, such as producing crisp spoken audio for movies, podcasts, or audiobooks. In the future this might be useful for real-time audio translation.

Work by @lematt1991 @apoorv2904 @adiyossLC @mhnt1580