At their developer conference this week, Microsoft announced that they're offering businesses their own private instances of GPT-4. They're also building something called Microsoft Graph that can...
At their developer conference this week, Microsoft announced that they're offering businesses their own private instances of GPT-4. They're also building something called Microsoft Graph that can ingest business documents, index them for the AI, and control access. They're also rolling out a wave of tools to make it easy to build AI chat apps for business. Microsoft says this is the future of work and that within 3 years, any data system or app that you can't have a conversation with will be "considered broken or invisible".
Ask your business intelligence program to generate a dashboard, and get one with animated graphs
Ask your email inbox which emails are most important
Automatically generate emails, Word documents, and PowerPoint presentations from prompts, pulling in other documents to supply data. ("Turn this report into a PowerPoint presentation.") It can also connect to external AI-based services; they give an example of calling various law plugins while drafting a contract to make sure the language is bulletproof.
Microsoft Teams is getting a Copilot bot that can read all the documents you have access to. It can hook into services like Jira and execute commands based on your instructions. Microsoft is making it really easy to make any API work with Copilot--your API just needs to be documented in the industry-standard OpenAPI format, and to have a JSON file that describes what your API does in English. Copilot will figure the rest out itself. I'm going to talk with my team this week about getting this in place for our internal APIs.
Not sure if this is the appropriate thread, but I see AI art supporters making the argument that a human studying existing art to learn how to make art of their own is not meaningfully different...
Not sure if this is the appropriate thread, but I see AI art supporters making the argument that a human studying existing art to learn how to make art of their own is not meaningfully different than the process by which AI image generators use existing art, and therefore concerns over the unauthorized use of existing images in the creation of an AI dataset are unwarranted or unfounded. How valid is this argument?
The argument is sound, but it is also not as relevant as it seems. IMHO, the main discussion here is not about what is art, but rather what is favorable, from an economic, societal, and subjective...
The argument is sound, but it is also not as relevant as it seems. IMHO, the main discussion here is not about what is art, but rather what is favorable, from an economic, societal, and subjective perspective.
In other words, questions regarding the ability to earn a livelihood, social recognition, as well as the subjective relationship humans have with their craft are, in my view, the ones we should focus on right now.
Can you elaborate? For instance are there specific mechanisms in the learning models that are similar to the human learning process? It's relevant in the sense that it's being used as a defense...
The argument is sound
Can you elaborate? For instance are there specific mechanisms in the learning models that are similar to the human learning process?
it is also not as relevant as it seems
It's relevant in the sense that it's being used as a defense against the specific criticism of using existing art without permission. Larger societal issues certainly exist but I'm making no attempt to address, or even worry about, them in the context of this one very specific idea I see being espoused by AI proponents.
I read a comment on here a few months ago that I thought laid it out extremely well. I was going to quote a snippet or summarise a bit, but honestly the top post on that thread is so spot on for...
I read a comment on here a few months ago that I thought laid it out extremely well. I was going to quote a snippet or summarise a bit, but honestly the top post on that thread is so spot on for your question that I’d say just give it a read.
As for larger issues, I’m not sure they can be separated, at least not so easily. Even as someone who finds the whole field fascinating, I can totally understand anxiety, negativity, even fear around these tools: they’re complex and likely to have far reaching consequences that we’re just scratching the surface of, just as the internet itself has done in the 30 years before. There are plenty of totally valid concerns to discuss here and now - but I’m also seeing a lot of repeated themes that are flat wrong and logically inconsistent being used as headlines against ML text and image generators. I don’t think these are necessarily made in deliberate bad faith most of the time, but I do think a lot of people are worried and jumping to arguments that counter that worry even when the points don’t really add up.
I expect that arguments about what is fair will go on for a while. I think, practically speaking, what is legal is what matters. There are court cases, but we don't know how they will turn out....
I expect that arguments about what is fair will go on for a while. I think, practically speaking, what is legal is what matters. There are court cases, but we don't know how they will turn out.
This isn't directly related, but the US copyright office has said that the output of an AI image generator can't be copyright protected. The results of that will be interesting if it holds up. Seems like some artists who use an AI image generator for work might have incentive to keep it secret?
In the meantime, people say stuff. Companies making AI generators keep doing whatever they're doing, and aren't always very forthcoming about where they got their source material.
That is reasonable. However, law does not exist in isolation, nor it should be. Prevalent ideas on what is fair and equitable are consequential to both legislative acts and jurisprudence....
That is reasonable. However, law does not exist in isolation, nor it should be. Prevalent ideas on what is fair and equitable are consequential to both legislative acts and jurisprudence. Especially when new phenomena arise, forcing institutions to come up with new classifications, concepts, rules, etc.
Yes it does, but how much what we say on the Internet indirectly influences a judge or a legislature is unknown. My guess is that it's usually a pretty negligible effect, and I prefer not to think...
Yes it does, but how much what we say on the Internet indirectly influences a judge or a legislature is unknown. My guess is that it's usually a pretty negligible effect, and I prefer not to think about it. It seems like assuming we're somehow influential has a bad effect on conversations?
I never assume I'm even remotely influential to anything. And yet, somehow I am. I'm influencing you a little bit, and you are influencing me. But I don't wanna get tangled in a consequentialist...
I never assume I'm even remotely influential to anything. And yet, somehow I am. I'm influencing you a little bit, and you are influencing me.
But I don't wanna get tangled in a consequentialist maze. I talk about these matters for the same reason I vote. I know I won't decide any election, but there is virtue in fulfilling a civic duty. This is meaningful to me.
Influencing the person you're actually having a conversation with seems like a fine reason to write. It's different from playing to the crowd in a situation where we can't see the crowd and have...
Influencing the person you're actually having a conversation with seems like a fine reason to write. It's different from playing to the crowd in a situation where we can't see the crowd and have to imagine it, or imagining even more indirect and unlikely influences.
We shouldn't forget that this is all public on the Internet. But an abstract duty to political discussion seems like part of the consequentialist maze? Suppose we assume for a moment that there is no crowd, no duty to write, and whatever you write is mostly for the person you're replying to? Is that good enough motivation to continue?
Though, I am often guilty of not reading carefully and being more interested in expressing my own ideas. I guess that's a reason to write, too. I find it hard to find the motivation to write without someone specific to write to, though.
Maybe political discussion on Tildes isn't for some abstract person out there who we imagine might benefit, it's just for us. Just as in a classroom, the discussion is primarily for the benefit of the students. If it's not working for us, we can stop.
The duty for civil engagement can be expressed in consequentialist terms, but this is not a very persuasive application of consequentialism in my view. The civic duty I talk about is rooted in...
But an abstract duty to political discussion seems like part of the consequentialist maze?
The duty for civil engagement can be expressed in consequentialist terms, but this is not a very persuasive application of consequentialism in my view. The civic duty I talk about is rooted in virtue ethics, which is pretty much in opposition to consequentialism. In virtue ethics, something is good due to its inherent moral character. Not its consequences.
Some online rationalists seem to believe they are fully consequentialists, but I don't think that is feasible. Some measure of virtue ethics is apparently essential to the sanity of most humans.
Well, the differences between different kinds of ethics can be fuzzy, making them not as opposed as you might think. For example, someone might believe that having moral rules are good for...
Well, the differences between different kinds of ethics can be fuzzy, making them not as opposed as you might think. For example, someone might believe that having moral rules are good for consequentialist reasons, which is called rule utilitarianism. And one reason they might think that is because it's very difficult and impractical to calculate the consequences in each case.
This is pretty similar to arguing that there should be a law based on what you think its benefits will be. Other people might be in favor of the same law because they believe it's morally right.
Also, some kinds of virtue ethics allow for exceptions for essentially consequentialist reasons.
You are absolutely correct that there are many subtypes and gradations philosophers use to categorize their ethics. Still, in a charitable interpretation, it is possible to understand that I'm...
You are absolutely correct that there are many subtypes and gradations philosophers use to categorize their ethics. Still, in a charitable interpretation, it is possible to understand that I'm talking about overall tendencies that are identifiable in discourse, and a tension between mindsets that are assumed to be almost entirely consequential with no qualification, and the reality that this is not the case.
Sure, fair enough. To try to get closer to the original discussion, what I understand you saying is that you don't justify your beliefs about civic duty or about participating in political...
Sure, fair enough. To try to get closer to the original discussion, what I understand you saying is that you don't justify your beliefs about civic duty or about participating in political discussions in consequentialist terms?
I don't think the time I spend online makes sense in consequentialist terms (it's more of a habit) but I do try to write about concrete benefits of doing things because I assume that's how persuasion works, if it works at all.
Yes, that is entirely correct. This is most often correct, but I find it valuable to note that not all ethics assessments can be sufficiently elucidated through consequence. Consequence seems to...
To try to get closer to the original discussion, what I understand you saying is that you don't justify your beliefs about civic duty or about participating in political discussions in consequentialist terms?
Yes, that is entirely correct.
I don't think the time I spend online makes sense in consequentialist terms (it's more of a habit) but I do try to write about concrete benefits of doing things because I assume that's how persuasion works, if it works at all
This is most often correct, but I find it valuable to note that not all ethics assessments can be sufficiently elucidated through consequence. Consequence seems to be an implicit default even when its shortcomings are most apparent.
Getting back to what started this digression, my claim was that what we say here doesn't matter much, it's just for us. I know you disagree with that, but it also means that if people want to...
Getting back to what started this digression, my claim was that what we say here doesn't matter much, it's just for us. I know you disagree with that, but it also means that if people want to discuss what's fair for AI-generated imagery, that's fine, too, and I really shouldn't object. Even if it turns out we have no real control over what happens, it could still be a good discussion for other reasons.
Well than I guess I'll surprise you because in a sense I agree with everything you say. Our conversations doesn't matter in the grand scheme of things but I hope it matters to you because it...
Well than I guess I'll surprise you because in a sense I agree with everything you say.
Our conversations doesn't matter in the grand scheme of things but I hope it matters to you because it certainly matters to me!
All the Hard Stuff Nobody Talks About when Building Products with LLMs (honeycomb.io) Apparently lots of things look like they work in demos, but if you try to build a product, they don't actually...
An overview describing what researchers hope to learn from mechanistic interpretability research.
The success of deep learning is often bemoaned as scientifically boring. One just stacks on layers, makes the models bigger, and gets lower loss. Elegant and clever ideas are often unnecessary. Sutton's essay The Bitter Lesson famously presents this as a painful lesson where many of one's beautiful ideas become unimportant as we throw more compute at machine learning.
We take the opposite view. The beauty of deep learning and scale is a kind of biological beauty. Just as the simplicity of evolution gives rise to incredible beauty and complexity in the natural world, so too does the simplicity of gradient descent on neural networks give rise to incredible beauty in machine learning.
Microsoft has announced Windows Copilot, a sidebar for Windows where you can enter natural-language commands for the computer to execute, like on Star Trek but typed instead of verbal. This looks...
Microsoft has announced Windows Copilot, a sidebar for Windows where you can enter natural-language commands for the computer to execute, like on Star Trek but typed instead of verbal. This looks like what I always wanted Cortana to be.
It can explain things that are on your screen. You can highlight some unfamiliar code and ask it for a line-by-line explanation, and it will do it. You can record an audio file with your microphone, and Copilot will transcribe it for you. You can hand it a PDF, and it will provide an explanation or a summary. This will be beta-tested by Windows Insiders in June.
OpenAI says it could ‘cease operating’ in the EU if it can’t comply with future regulation (The Verge) Note that Google’s Bard isn’t available in most of Europe or Canada, though it is available...
Speaking to reporters after a talk in London, Altman said he had “many concerns” about the EU AI Act, which is currently being finalized by lawmakers. The terms of the Act have been expanded in recent months to include new obligations for makers of so-called “foundation models” — large-scale AI systems that power services like OpenAI’s ChatGPT and DALL-E.
“The details really matter,” said Altman, according to a report from The Financial Times. “We will try to comply, but if we can’t comply we will cease operating.”
In comments reported by Time, Altman said the concern was that systems like ChatGPT would be designated “high risk” under the EU legislation. This means OpenAI would have to meet a number of safety and transparency requirements. “Either we’ll be able to solve those requirements or not,” said Altman. “[T]here are technical limits to what’s possible.”
Note that Google’s Bard isn’t available in most of Europe or Canada, though it is available in the UK. Perhaps Google is worried about regulation too?
I’m wondering if regulatory uncertainty would end up being a boost to open source AI projects? Or maybe people will learn to use VPN’s more?
Scientists using artificial intelligence have discovered a new antibiotic that can kill a deadly superbug.
According to a new study published on Thursday in the science journal Nature Chemical Biology, a group of scientists from McMaster University and the Massachusetts Institute of Technology have discovered a new antibiotic that can be used to kill a deadly hospital superbug.
The superbug in question is Acinetobacter baumannii, which the World Health Organization has classified as a “critical” threat among its “priority pathogens” – a group of bacteria families that pose the “greatest threat” to human health.
This is AI but not a language model or a generative model of the type used to make images. Apparently it's a "message-passing deep neural network," also called a "graph neural network." I found A...
This is AI but not a language model or a generative model of the type used to make images. Apparently it's a "message-passing deep neural network," also called a "graph neural network." I found A Gentle Introduction to Graph Neural Networks, which looks helpful.
A lawyer asked ChatGPT for examples of cases that supported an argument they were trying to make.
ChatGPT, as it often does, hallucinated wildly—it invented several supporting cases out of thin air.
When the lawyer was asked to provide copies of the cases in question, they turned to ChatGPT for help again—and it invented full details of those cases, which they duly screenshotted and copied into their legal filings.
At some point, they asked ChatGPT to confirm that the cases were real... and ChatGPT said that they were. They included screenshots of this in another filing.
The judge is furious. Many of the parties involved are about to have a very bad time.
It turns out that LLMs are very good at imitation! Who knew? :-) Here's a paper about how many smaller LLM's aren't as good as they appear. Apparently if you train a small LLM to imitate a large...
It turns out that LLMs are very good at imitation! Who knew? :-)
Here's a paper about how many smaller LLM's aren't as good as they appear. Apparently if you train a small LLM to imitate a large one, crowd-sourced evaluators think its output is as good as the big one about 70% of the time, but this is an illusion.
I expect that the ease of making demos that seem to work will probably result in a lot of low-quality AI chat apps. This doesn't seem like good news for consumers or for venture capitalists easily fooled by a good demo.
I haven't bothered to try any of them yet. Not sure why I should, unless I hear about something that a lot of people say is better than what I have easy access to?
An emerging method to cheaply improve a weaker language model is to finetune it on outputs from a stronger model, such as a proprietary system like ChatGPT (e.g., Alpaca, Self-Instruct, and others). This approach looks to cheaply imitate the proprietary model's capabilities using a weaker open-source model. In this work, we critically analyze this approach. We first finetune a series of LMs that imitate ChatGPT using varying base model sizes (1.5B--13B), data sources, and imitation data amounts (0.3M--150M tokens). We then evaluate the models using crowd raters and canonical NLP benchmarks. Initially, we were surprised by the output quality of our imitation models -- they appear far better at following instructions, and crowd workers rate their outputs as competitive with ChatGPT. However, when conducting more targeted automatic evaluations, we find that imitation models close little to none of the gap from the base LM to ChatGPT on tasks that are not heavily supported in the imitation data. We show that these performance discrepancies may slip past human raters because imitation models are adept at mimicking ChatGPT's style but not its factuality. Overall, we conclude that model imitation is a false promise: there exists a substantial capabilities gap between open and closed LMs that, with current methods, can only be bridged using an unwieldy amount of imitation data or by using more capable base LMs. In turn, we argue that the highest leverage action for improving open-source models is to tackle the difficult challenge of developing better base LMs, rather than taking the shortcut of imitating proprietary systems.
Statement on AI Risk It's very short: A bunch of famous people and academics have signed it. So I guess this is people saying that AI doom is (or should be) a live issue, in a way so that it can't...
Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.
A bunch of famous people and academics have signed it.
So I guess this is people saying that AI doom is (or should be) a live issue, in a way so that it can't be dismissed as just a crackpot thing. It doesn't get us much closer to agreement on doing anything, though.
Google Bard now includes images in its responses This means it can do an image search. It doesn't generate images. You can ask it questions about the images, but it seems it can't actually see...
This means it can do an image search. It doesn't generate images.
You can ask it questions about the images, but it seems it can't actually see them. I asked it for parrot images, and then asked which one is a blue parrot. It showed one of the images that's from the Wikipedia page for a blue and yellow macaw, but not the one of the hyacinth macaw, which is almost entirely blue.
Geoffrey Hinton was one of the loudest voices decrying the decline of radiology 5 years, and now he’s crying fear for new AI systems.
There's a lot to unpack for both why Geof was wrong, and why his future predictions should not be taken seriously either. Geoff made a classic error that technologists often make, which is to observe a particular behavior (identifying some subset of radiology scans correctly) against some task (identifying hemorrhage on CT head scans correctly), and then to extrapolate based on that task alone.
The reality is that reducing any job, especially a wildly complex job that requires a decade of training, to a handful of tasks is quite absurd. Here's a bunch of stuff you wouldn't know about radiologists unless you built an AI company with them [...]
A minor but nice improvement: OpenAI added a way to share a link to a ChatGPT conversation. (Before, people would use a browser extension to do this.) You can also continue someone else's...
A minor but nice improvement: OpenAI added a way to share a link to a ChatGPT conversation. (Before, people would use a browser extension to do this.) You can also continue someone else's conversation.
I tried it and it mostly worked, but at the top of the page it says "Model: Default" instead of "Model: GPT-4" which is what I used. When I continued the conversation it used GPT-4. I was hoping it would let you switch models.
It looks like AI chat will someday be able to see charts and graphs and answer questions about them: Foundation models for reasoning on charts (Google AI Blog) They released the code on GitHub and...
It looks like AI chat will someday be able to see charts and graphs and answer questions about them:
[W]e propose “MatCha: Enhancing Visual Language Pretraining with Math Reasoning and Chart Derendering”. MatCha, which stands for math and charts, is a pixels-to-text foundation model (a pre-trained model with built-in inductive biases that can be fine-tuned for multiple applications) trained on two complementary tasks: (a) chart de-rendering and (b) math reasoning. In chart de-rendering, given a plot or chart, the image-to-text model is required to generate its underlying data table or the code used to render it. For math reasoning pre-training, we pick textual numerical reasoning datasets and render the input into images, which the image-to-text model needs to decode for answers. We also propose “DePlot: One-shot visual language reasoning by plot-to-table translation”, a model built on top of MatCha for one-shot reasoning on charts via translation to tables. With these methods we surpass the previous state of the art in ChartQA by more than 20% and match the best summarization systems that have 1000 times more parameters.
They released the code on GitHub and apparently you can try it in Google Colab. (Not something I've used.)
At their developer conference this week, Microsoft announced that they're offering businesses their own private instances of GPT-4. They're also building something called Microsoft Graph that can ingest business documents, index them for the AI, and control access. They're also rolling out a wave of tools to make it easy to build AI chat apps for business. Microsoft says this is the future of work and that within 3 years, any data system or app that you can't have a conversation with will be "considered broken or invisible".
They're integrating GPT into all the Office apps, with features like:
Microsoft Teams is getting a Copilot bot that can read all the documents you have access to. It can hook into services like Jira and execute commands based on your instructions. Microsoft is making it really easy to make any API work with Copilot--your API just needs to be documented in the industry-standard OpenAPI format, and to have a JSON file that describes what your API does in English. Copilot will figure the rest out itself. I'm going to talk with my team this week about getting this in place for our internal APIs.
Not sure if this is the appropriate thread, but I see AI art supporters making the argument that a human studying existing art to learn how to make art of their own is not meaningfully different than the process by which AI image generators use existing art, and therefore concerns over the unauthorized use of existing images in the creation of an AI dataset are unwarranted or unfounded. How valid is this argument?
The argument is sound, but it is also not as relevant as it seems. IMHO, the main discussion here is not about what is art, but rather what is favorable, from an economic, societal, and subjective perspective.
In other words, questions regarding the ability to earn a livelihood, social recognition, as well as the subjective relationship humans have with their craft are, in my view, the ones we should focus on right now.
Is it fair? is more relevant than is it art?.
You might say I'm very Marxist on that subject.
Can you elaborate? For instance are there specific mechanisms in the learning models that are similar to the human learning process?
It's relevant in the sense that it's being used as a defense against the specific criticism of using existing art without permission. Larger societal issues certainly exist but I'm making no attempt to address, or even worry about, them in the context of this one very specific idea I see being espoused by AI proponents.
I read a comment on here a few months ago that I thought laid it out extremely well. I was going to quote a snippet or summarise a bit, but honestly the top post on that thread is so spot on for your question that I’d say just give it a read.
As for larger issues, I’m not sure they can be separated, at least not so easily. Even as someone who finds the whole field fascinating, I can totally understand anxiety, negativity, even fear around these tools: they’re complex and likely to have far reaching consequences that we’re just scratching the surface of, just as the internet itself has done in the 30 years before. There are plenty of totally valid concerns to discuss here and now - but I’m also seeing a lot of repeated themes that are flat wrong and logically inconsistent being used as headlines against ML text and image generators. I don’t think these are necessarily made in deliberate bad faith most of the time, but I do think a lot of people are worried and jumping to arguments that counter that worry even when the points don’t really add up.
I expect that arguments about what is fair will go on for a while. I think, practically speaking, what is legal is what matters. There are court cases, but we don't know how they will turn out.
This isn't directly related, but the US copyright office has said that the output of an AI image generator can't be copyright protected. The results of that will be interesting if it holds up. Seems like some artists who use an AI image generator for work might have incentive to keep it secret?
In the meantime, people say stuff. Companies making AI generators keep doing whatever they're doing, and aren't always very forthcoming about where they got their source material.
That is reasonable. However, law does not exist in isolation, nor it should be. Prevalent ideas on what is fair and equitable are consequential to both legislative acts and jurisprudence. Especially when new phenomena arise, forcing institutions to come up with new classifications, concepts, rules, etc.
Yes it does, but how much what we say on the Internet indirectly influences a judge or a legislature is unknown. My guess is that it's usually a pretty negligible effect, and I prefer not to think about it. It seems like assuming we're somehow influential has a bad effect on conversations?
I never assume I'm even remotely influential to anything. And yet, somehow I am. I'm influencing you a little bit, and you are influencing me.
But I don't wanna get tangled in a consequentialist maze. I talk about these matters for the same reason I vote. I know I won't decide any election, but there is virtue in fulfilling a civic duty. This is meaningful to me.
Otherwise, why engage at all?
Influencing the person you're actually having a conversation with seems like a fine reason to write. It's different from playing to the crowd in a situation where we can't see the crowd and have to imagine it, or imagining even more indirect and unlikely influences.
We shouldn't forget that this is all public on the Internet. But an abstract duty to political discussion seems like part of the consequentialist maze? Suppose we assume for a moment that there is no crowd, no duty to write, and whatever you write is mostly for the person you're replying to? Is that good enough motivation to continue?
Though, I am often guilty of not reading carefully and being more interested in expressing my own ideas. I guess that's a reason to write, too. I find it hard to find the motivation to write without someone specific to write to, though.
Maybe political discussion on Tildes isn't for some abstract person out there who we imagine might benefit, it's just for us. Just as in a classroom, the discussion is primarily for the benefit of the students. If it's not working for us, we can stop.
The duty for civil engagement can be expressed in consequentialist terms, but this is not a very persuasive application of consequentialism in my view. The civic duty I talk about is rooted in virtue ethics, which is pretty much in opposition to consequentialism. In virtue ethics, something is good due to its inherent moral character. Not its consequences.
Some online rationalists seem to believe they are fully consequentialists, but I don't think that is feasible. Some measure of virtue ethics is apparently essential to the sanity of most humans.
Well, the differences between different kinds of ethics can be fuzzy, making them not as opposed as you might think. For example, someone might believe that having moral rules are good for consequentialist reasons, which is called rule utilitarianism. And one reason they might think that is because it's very difficult and impractical to calculate the consequences in each case.
This is pretty similar to arguing that there should be a law based on what you think its benefits will be. Other people might be in favor of the same law because they believe it's morally right.
Also, some kinds of virtue ethics allow for exceptions for essentially consequentialist reasons.
You are absolutely correct that there are many subtypes and gradations philosophers use to categorize their ethics. Still, in a charitable interpretation, it is possible to understand that I'm talking about overall tendencies that are identifiable in discourse, and a tension between mindsets that are assumed to be almost entirely consequential with no qualification, and the reality that this is not the case.
Sure, fair enough. To try to get closer to the original discussion, what I understand you saying is that you don't justify your beliefs about civic duty or about participating in political discussions in consequentialist terms?
I don't think the time I spend online makes sense in consequentialist terms (it's more of a habit) but I do try to write about concrete benefits of doing things because I assume that's how persuasion works, if it works at all.
Yes, that is entirely correct.
This is most often correct, but I find it valuable to note that not all ethics assessments can be sufficiently elucidated through consequence. Consequence seems to be an implicit default even when its shortcomings are most apparent.
Getting back to what started this digression, my claim was that what we say here doesn't matter much, it's just for us. I know you disagree with that, but it also means that if people want to discuss what's fair for AI-generated imagery, that's fine, too, and I really shouldn't object. Even if it turns out we have no real control over what happens, it could still be a good discussion for other reasons.
Well than I guess I'll surprise you because in a sense I agree with everything you say.
Our conversations doesn't matter in the grand scheme of things but I hope it matters to you because it certainly matters to me!
All the Hard Stuff Nobody Talks About when Building Products with LLMs (honeycomb.io)
Apparently lots of things look like they work in demos, but if you try to build a product, they don't actually work very well.
(Via Simon Willison; I'm getting a lot of good links from his blog these days.)
Just to lighten the mood: I'm now the proud owner of GAIagenda.com and there will be GAI shenanigans. 🌈
Interpretability Dreams (Chris Olah)
An overview describing what researchers hope to learn from mechanistic interpretability research.
Microsoft has announced Windows Copilot, a sidebar for Windows where you can enter natural-language commands for the computer to execute, like on Star Trek but typed instead of verbal. This looks like what I always wanted Cortana to be.
It can explain things that are on your screen. You can highlight some unfamiliar code and ask it for a line-by-line explanation, and it will do it. You can record an audio file with your microphone, and Copilot will transcribe it for you. You can hand it a PDF, and it will provide an explanation or a summary. This will be beta-tested by Windows Insiders in June.
OpenAI says it could ‘cease operating’ in the EU if it can’t comply with future regulation (The Verge)
Note that Google’s Bard isn’t available in most of Europe or Canada, though it is available in the UK. Perhaps Google is worried about regulation too?
I’m wondering if regulatory uncertainty would end up being a boost to open source AI projects? Or maybe people will learn to use VPN’s more?
A day after he called for regulation in the US Congress…
https://www.axios.com/2023/05/17/ai-leaders-sam-altman-regulate-senate
Scientists use AI to discover new antibiotic to treat deadly superbug
The Guardian – Maya Yang – 25th May 2023
This is AI but not a language model or a generative model of the type used to make images. Apparently it's a "message-passing deep neural network," also called a "graph neural network." I found A Gentle Introduction to Graph Neural Networks, which looks helpful.
Lawyer cites fake cases invented by ChatGPT, judge is not amused (Simon Willison)
The blog post has a detailed timeline.
It turns out that LLMs are very good at imitation! Who knew? :-)
Here's a paper about how many smaller LLM's aren't as good as they appear. Apparently if you train a small LLM to imitate a large one, crowd-sourced evaluators think its output is as good as the big one about 70% of the time, but this is an illusion.
I expect that the ease of making demos that seem to work will probably result in a lot of low-quality AI chat apps. This doesn't seem like good news for consumers or for venture capitalists easily fooled by a good demo.
I haven't bothered to try any of them yet. Not sure why I should, unless I hear about something that a lot of people say is better than what I have easy access to?
The False Promise of Imitating Proprietary LLMs
Statement on AI Risk
It's very short:
A bunch of famous people and academics have signed it.
So I guess this is people saying that AI doom is (or should be) a live issue, in a way so that it can't be dismissed as just a crackpot thing. It doesn't get us much closer to agreement on doing anything, though.
Google Bard now includes images in its responses
This means it can do an image search. It doesn't generate images.
You can ask it questions about the images, but it seems it can't actually see them. I asked it for parrot images, and then asked which one is a blue parrot. It showed one of the images that's from the Wikipedia page for a blue and yellow macaw, but not the one of the hyacinth macaw, which is almost entirely blue.
Production AI systems are really hard (Kevin Fischer)
A minor but nice improvement: OpenAI added a way to share a link to a ChatGPT conversation. (Before, people would use a browser extension to do this.) You can also continue someone else's conversation.
I tried it and it mostly worked, but at the top of the page it says "Model: Default" instead of "Model: GPT-4" which is what I used. When I continued the conversation it used GPT-4. I was hoping it would let you switch models.
It looks like AI chat will someday be able to see charts and graphs and answer questions about them:
Foundation models for reasoning on charts (Google AI Blog)
They released the code on GitHub and apparently you can try it in Google Colab. (Not something I've used.)