• Yusuf Mehdi at Microsoft:

    Today, we’re launching an all new, AI-powered Bing search engine and Edge browser, available in preview now at Bing.com, to deliver better search, more complete answers, a new chat experience and the ability to generate content. We think of these tools as an AI copilot for the web.

    […]

    There are 10 billion search queries a day, but we estimate half of them go unanswered. That’s because people are using search to do things it wasn’t originally designed to do. It’s great for finding a website, but for more complex questions or tasks too often it falls short.

    […]

    We’re excited to announce the new Bing is running on a new, next-generation OpenAI large language model that is more powerful than ChatGPT and customized specifically for search. It takes key learnings and advancements from ChatGPT and GPT-3.5 – and it is even faster, more accurate and more capable.

    The most striking aspect of this whole generative AI saga is that it has put Google in the position of playing catch-up. Even if their products end up being superior to Microsoft’s — which is a distinct possibility — the narrative will be that Bard is a response to ChatGPT and their chat-based search features are a copy of “the new Bing.” You can’t rest on your laurels forever, I guess.

  • First, Google’s response to ChatGPT — Bard

    Sundar Pichai:

    We’ve been working on an experimental conversational AI service, powered by LaMDA, that we’re calling Bard. And today, we’re taking another step forward by opening it up to trusted testers ahead of making it more widely available to the public in the coming weeks.

    Bard seeks to combine the breadth of the world’s knowledge with the power, intelligence and creativity of our large language models. It draws on information from the web to provide fresh, high-quality responses.

    The fact that Bard will include information from the web might make this a really big deal. Now we will just have to wait and see what kind of guardrails Google puts on this: if Bard is more restrictive than ChatGPT then it is quite possible none of these other improvements will matter. Regardless, we are in for some fascinating competition.

    Now, here’s the big one: search

    Increasingly, people are turning to Google for deeper insights and understanding — like, “is the piano or guitar easier to learn, and how much practice does each need?” Learning about a topic like this can take a lot of effort to figure out what you really need to know, and people often want to explore a diverse range of opinions or perspectives.

    AI can be helpful in these moments, synthesizing insights for questions where there’s no one right answer. Soon, you’ll see AI-powered features in Search that distill complex information and multiple perspectives into easy-to-digest formats, so you can quickly understand the big picture and learn more from the web.

    I can’t wait to try this myself. It does, however, look like they are still taking a slow, cautious approach to rolling out generative AI features into Search. Notice how, in the quote above, they mention integrating AI insights into “questions where there’s no one right answer.”

  • Lars Doucet writes about his predictions for the near-term future of the internet:

    What happens when anyone can spin up a thousand social media accounts at the click of a button, where each account picks a consistent persona and sticks to it – happily posting away about one of their hobbies like knitting or trout fishing or whatever, while simultaneously building up a credible and inobtrusive post history in another plausible side hobby that all these accounts happen to share – geopolitics, let’s say – all until it’s time for the sock puppet master to light the bat signal and manufacture some consensus?

    What happens when most “people” you interact with on the internet are fake?

  • Jennifer Elias, reporting for CNBC:

    Google is testing new artificial intelligence-powered chat products that are likely to influence a future public product launch.

    […]

    One of the test products is a chatbot called Apprentice Bard, which uses Google’s conversation technology LaMDA… Employees can enter a question in a dialog box and get a text answer, then give feedback on the response. Based on several responses viewed by CNBC, Apprentice Bard’s answers can include recent events, a feature ChatGPT doesn’t have yet.

    […]

    The company is also testing an alternate search page that could use a question-and-answer format, according to designs viewed by CNBC… When a question is entered, the search results show a gray bubble directly under the search bar, offering more human-like responses than typical search results. Directly beneath that, the page suggests several follow-up questions related to the first one. Under that, it shows typical search results, including links and headlines.

    The potential new search page sounds pretty similar to what I described as an ideal interface last month. Microsoft on the other hand…

    James Vincent at The Verge:

    Student and designer Owen Yin reported seeing the “new Bing” on Twitter this morning.

    […]

    Screenshots of the AI-augmented Bing show a new “chat” option appearing in the menu bar next to “search.” Select it and you’re taken to a chat interface that says, “Welcome to the new Bing: Your AI-powered answer engine.”

    Definitely visit the story above to see (alleged) screenshots of the new Bing interface. I am going to try to withhold judgment until this feature is officially released but at the moment it looks like, instead of actually working to create any kind of meaningful search and chat integration, Microsoft just slapped a ChatGPT tab onto the bing.com homepage. I hope they continue iterating before making this public.

  • Ingrid Lunden, reporting for TechCrunch:

    Customers of Shutterstock’s Creative Flow online design platform will now be able to create images based on text prompts, powered by OpenAI and Dall-E 2… Shutterstock says the images are “ready for licensing” right after they’re made.

    As far as I can tell, you are already free to use images created with Dall-E 2 commercially so I am not entirely sure how the Stutterstock partnership changes things. It does, however present a stark contrast when compared to Getty Images’ response to generative AI. The article continues:

    One of Shutterstock’s big competitors, Getty Images, is currently embroiled in a lawsuit against Stability AI — maker of another generative AI service called Stable Diffusion — over using its images to train its AI without permission from Getty or rightsholders.

    In other words, Shutterstock’s service is not only embracing the ability to use AI… but it’s setting the company up in opposition to Getty in terms of how it is embracing the brave new world of artificial intelligence.

    My knee-jerk reaction is to say that Getty is behind the times here but, after thinking about this a little bit more, I am less sure about that.

    If Shutterstock starts re-licensing AI generated images, why would you pay for them instead paying of OpenAI or Midjourney directly? More to the point, why not use Stable Diffusion to generate images, for free, on your own computer?

    Getty Images, on the other hand, gets to be the anti-AI company selling certified human-made images. I can see that being a valuable niche for some time to come.

  • It turns out, Google published a research paper back in 2021 detailing how an AI-based information retrieval system could be an alternative to traditional search engines. This predates ChatGPT and search services build on top of it like Perplexity.ai. The paper begins with some background on why this would be a worthwhile development:

    Given an information need, users often turn to search engines for help. Such systems point them in the direction of one or more relevant items from a corpus. This is appropriate for navigational and transactional intents (e.g. home page finding or online shopping) but typically less ideal for informational needs, where users seek answers to questions they may have… The very fact that ranking is a critical component of this paradigm is a symptom of the retrieval system providing users a selection of potential answers, which induces a rather significant cognitive burden on the user.

    […]

    State-of-the-art pre-trained language models are capable of directly generating prose that may be responsive to an information need. However, such models are dilettantes – they do not have a true understanding of the world, they are prone to hallucinating, and crucially they are incapable of justifying their utterances by referring to supporting documents in the corpus they were trained over.

    Taking a step back, it is true that search engines, like Google, are best for retrieving a list of relevant documents, not answering questions. This is intuitive but easily forgotten among all of the hype around systems like ChatGPT.

    The paper goes on to list some areas in need of further research:

    • How to implement continuous, incremental learning
    • How to make the AI “forget” specific pieces of information, for legal or privacy reasons
    • How to ensure the model is predictable, interpretable, and debuggable
    • How to lower inference costs (i.e. the cost to run each query)

    With all of the fervor around AI-based search, it will be interesting to see how many of these points are still open problems a year from now.


    In related news, Reed Albergotti at Semafor reported that GPT-4 might appear in Bing soon?

    Microsoft’s second-place search engine Bing is poised to incorporate a faster and richer version of ChatGPT, known as GPT-4, into its product in the coming weeks

    […]

    OpenAI is also planning to launch a mobile ChatGPT app and test a new feature in its Dall-E image-generating software that would create videos with the help of artificial intelligence.

    […]

    The most interesting improvement in the latest version described by sources is GPT-4’s speed

    This is such a strange set of rumors; especially the fact that the only noted change in GPT-4 is it’s speed. Plus, it is being launched as a Bing integration? A new version of Dall-E for video would be super exciting; I am really skeptical about everything else in this report.

  • From OpenAI’s announcement:

    We’ve trained a classifier to distinguish between text written by a human and text written by AIs from a variety of providers. While it is impossible to reliably detect all AI-written text, we believe good classifiers can inform mitigations for false claims that AI-generated text was written by a human

    The classifier is accessible for free on OpenAI’s website. My apologies to Turnitin and GPTZero.

    Near the end of the announcement, OpenAI links to their new Educator considerations for ChatGPT page.

    We recognize that many school districts and higher education institutions do not currently account for generative AI in their policies on academic dishonesty. We also understand that many students have used these tools for assignments without disclosing their use of AI. Each institution will address these gaps in a way and on a timeline that makes sense for their educators and students. We do however caution taking punitive measures against students for using these technologies if proper expectations are not set ahead of time for what users are or are not allowed.

    Classifiers such as the OpenAI AI text classifier can be helpful in detecting AI-generated content, but they are far from foolproof. These tools will produce both false negatives, where they don’t identify AI-generated content as such, and false positives, where they flag human-written content as AI-generated. Additionally, students may quickly learn how to evade detection by modifying some words or clauses in generated content.

    Ultimately, we believe it will be necessary for students to learn how to navigate a world where tools like ChatGPT are commonplace… Some of this is STEM education, but much of it also draws on students’ understanding of ethics, media literacy, ability to verify information from different sources, and other skills from the arts, social sciences, and humanities.

    Indeed, we are in a transitionary period where tools like OpenAI’s classifier are necessary. The most important work, now, will be figuring out how to integrate generative AI into education in healthy and productive ways. That’s the exciting part, too.

  • Zheping Huang, reporting for Bloomberg:

    [Baidu Inc.] plans to debut a ChatGPT-style application in March, initially embedding it into its main search services… The tool, whose name hasn’t been decided, will allow users to get conversation-style search results much like OpenAI’s popular platform.

    […]

    Baidu has spent billions of dollars researching AI in a years-long effort to transition from online marketing to deeper technology. Its Ernie system — a large-scale machine-learning model that’s been trained on data over several years — will be the foundation of its upcoming ChatGPT-like tool

    See also: On censorship of LLM models:

    It’s quite hard to restrict the output of general purpose, generative, black box algorithms… Censoring via limiting the training data is hard because algorithms could synthesize an “offensive” output by combining multiple outputs that are ok on their own… Adding an extra filter layer to censor is hard as well Look at all the trouble chatGPT has had with this. Users have repeatedly found ways around the dumb limitations on certain topics.

    I don’t think censoring the output of a language model is impossible but it is a fundamentally different problem than China has previously faced with its “Great Firewall.” It will be fascinating to see how Baidu’s engineers end up approaching it and whether this will ultimately impede the growth of these technolgies in China.

  • Agostinelli et al. at Google Research:

    We introduce MusicLM, a model generating high-fidelity music from text descriptions such as “a calming violin melody backed by a distorted guitar riff”. MusicLM casts the process of conditional music generation as a hierarchical sequence-to-sequence modeling task, and it generates music at 24 kHz that remains consistent over several minutes.

    Definitely listen to these demos — especially the AI generated voices singing and rapping in artificial languages. Amazing.

    Riffusion is still my favorite text-to-music demo, mostly because of the unbelievable way it works. I am, of course, excited to see more development here, though. The output from MusicLM is clearly better than Riffusion; I just wish there was a public demo I could try out.

  • Connie Guglielmo at CNET:

    In November, one of our editorial teams, CNET Money, launched a test using an internally designed AI engine – not ChatGPT – to help editors create a set of basic explainers around financial services topics. We started small and published 77 short stories using the tool, about 1% of the total content published on our site during the same period. Editors generated the outlines for the stories first, then expanded, added to and edited the AI drafts before publishing. After one of the AI-assisted stories was cited, rightly, for factual errors, the CNET Money editorial team did a full audit.

    […]

    As always when we find errors, we’ve corrected these stories, with an editors' note explaining what was changed. We’ve paused and will restart using the AI tool when we feel confident the tool and our editorial processes will prevent both human and AI errors.

    77 is one percent of stories? So a normal November for CNET is to produce almost 8,000 pieces? No wonder they are looking for an AI to help.

    I don’t think there is anything fundamentally wrong about using LLMs as a tool to assist in the writing process. Something changes, though, when you start treating it as a full-blown writer—when you give it a byline. At that point I think you should start being a little more introspective about how important the work you’re producing really is.

    James Vincent, reporting for The Verge:

    BuzzFeed says it’s going to use AI tools provided by ChatGPT creator OpenAI to “enhance” and “personalize” its content, according to a memo sent this morning to staff by CEO Jonah Peretti

    […]

    “Our industry will expand beyond AI-powered curation (feeds), to AI-powered creation (content),” says Peretti. “AI opens up a new era of creativity, where creative humans like us play a key role providing the ideas, cultural currency, inspired prompts, IP, and formats that come to life using the newest technologies.”

    […]

    In an example cited by the WSJ but not included in the memo, AI could be used to generate personalized rom-com pitches for readers.

    […]

    When asked by The Verge if BuzzFeed was considering using AI in its newsroom, the company’s VP of communications, Matt Mittenthal, replied, “No.”

    There is no need to reject the use of new technologies; by all means, experiment! But I am worried using AI to create content out of whole cloth risks devaluing all of the work you produce. Instead, using AI for personalization and curation will be much healthier step forward. I think BuzzFeed is on the right track here. CNET, less so.

  • Bob Allyn, reporting for NPR:

    Joshua Browder, the CEO of the New York-based startup DoNotPay, created a way for people contesting traffic tickets to use arguments in court generated by artificial intelligence.

    Here’s how it was supposed to work: The person challenging a speeding ticket would wear smart glasses that both record court proceedings and dictate responses into the defendant’s ear from a small speaker. The system relied on a few leading AI text generators, including ChatGPT and DaVinci.

    […]

    “Multiple state bar associations have threatened us,” Browder said. “One even said a referral to the district attorney’s office and prosecution and prison time would be possible.”

    “Even if it wouldn’t happen, the threat of criminal charges was enough to give it up,” he said. “The letters have become so frequent that we thought it was just a distraction and that we should move on.”

    Although I don’t think it is especially smart to expect a large language model will offer you cogent legal advice I am surprised it is illegal. If I could, for example, research the relevant laws using Google, textbooks, etc and then represent myself in court why couldn’t I use ChatGPT to do the same? I guess the problem probably lies with DoNotPay trying to charge for this service.

    With that said, obviously I’m not a lawyer (I’m not even a large language model). So I am quite out of my depth here.

    Instead of trying to help those accused of traffic violations use AI in the courtroom, Browder said DoNotPay will train its focus on assisting people dealing with expensive medical bills, unwanted subscriptions and issues with credit reporting agencies.

    I saw Browder’s demonstration of using ChatGPT to negotiate Comcast bills a little while ago. It is pretty impressive and, more importantly, the stakes are much lower than when writing legal arguments. Besides, although ChatGPT can pass some law school exams, it is not quite ready to take the bar.

  • From a recent debate between Gary Marcus and Grady Booch on AGI timelines:

    Marcus:

    I get that AGI is hard, and that we aren’t there yet. I think we are wasting funding and bright young minds on an approach that probably isn’t on the right path… But I am cautiously optimistic that we’ll do better in the next 75 [years], that once the hype cools off, people will finally dive deeper into neurosymbolic AI, and start to take some important steps. Our data problems are solved, our compute problems are mostly solved; it’s now mostly a matter of software, and of rethinking how we build AI. Why be so sure we can’t do that in the next 75 years?

    Booch:

    You posit that we will see AGI within a few decades, I think it is more like a few generations… With every step we move forward, we discover things we did not know we needed to know. It took evolution about 300 million years to move from the first organic neurons to where we are today, and I don’t think we can compress the remaining software problems associated with AGI in the next few decades.

    Marcus:

    In my darkest moments, I actually agree with you. For one thing, most of the money right now is not going to the wrong place: it’s mostly going to large language models, and for you, like for me, that just seems like an approximation to intelligence, not the real thing… But I see some signs that are promising. The neurosymbolic AI community is growing fast; conferences that used to be dozens are now thousands… I take that as a hopeful sign that the scaling-über-alas narrative is losing force, and that more and more people are open to new things.

    […]

    The rubber-that-meets-the-road question in the end is how many key discoveries do still we need to make, and how long do we need to make them?

    Booch:

    Intelligence is, for me, just the first phase in a spectrum that collectively we might speak of as synthetic sentience. Intelligence, I think, encompasses reasoning and learning. Indeed, I think in the next few decades, we will see astonishing progress in how we can build software-intensive systems that attend to inductive, deductive, and abductive reasoning.

    […]

    Consciousness and self-consciousness are the next phases in my spectrum. I suspect we’ll see some breakthroughs in ways to represent long term and short term memory, in our ability to represent theories of the world, theories of others, and theories of the self.

    […]

    Sentience and then sapience fill out this spectrum. The world of AI has not made a lot of progress in the past several years, nor do I see much attention being spent here… Work needs to be done in the area of planning, decision making, goals and agency, and action selection. We also need to make considerable progress in metacognition and mechanisms for subjective experience.

    […]

    These things, collectively, define what I’d call a synthetic mind. In the next decade, we will likely make interesting progress in all those parts I mentioned. But, we still don’t know how to architect these parts into a whole… This is not a problem of scale; this is not a problem of hardware. This is a problem of architecture.

  • Confirming the recent rumors Microsoft and OpenAI officially announced that they are expanding their (already close) partnership. From Microsoft’s announcement:

    Today, we are announcing the third phase of our long-term partnership with OpenAI through a multiyear, multibillion dollar investment to accelerate AI breakthroughs to ensure these benefits are broadly shared with the world.

    This agreement follows our previous investments in 2019 and 2021. It extends our ongoing collaboration across AI supercomputing and research and enables each of us to independently commercialize the resulting advanced AI technologies.

    OpenAI’s announcement:

    In pursuit of our mission to ensure advanced AI benefits all of humanity, OpenAI remains a capped-profit company and is governed by the OpenAI non-profit. This structure allows us to raise the capital we need to fulfill our mission without sacrificing our core beliefs about broadly sharing benefits and the need to prioritize safety.

    Microsoft shares this vision and our values, and our partnership is instrumental to our progress.

    I am not sure Microsoft has received all of the credit they deserve from how strategically smart their OpenAI partnership is. Microsoft gets to reap the rewards from OpenAI’s advances while maintaining the ability to easily distance itself from the partnership in the event of any future controversies. Google doesn’t have that luxury and will therefore likely move much more slowly, fearing reputational risks if they get anything wrong.

  • Matthew Ball writes about why it can seem like AR/VR technology is perpetually “only a few years away” from mass adoption:

    As we observe the state of XR in 2023, it’s fair to say the technology has proved harder than many of the best-informed and most financially endowed companies expected. When it unveiled Google Glass, Google suggested that annual sales could reach the tens of millions by 2015, with the goal of appealing to the nearly 80% of people who wear glasses daily. Though Google continues to build AR devices, Glass was an infamous flop, with sales in the tens of thousands.

    […]

    Throughout 2015 and 2016, Mark Zuckerberg repeated his belief that within a decade, “normal-looking” AR glasses might be a part of daily life. Now it looks like Facebook won’t launch a dedicated AR headset by 2025—let alone an edition that hundreds of millions might want.

    […]

    In 2016, Epic Games founder/CEO Tim Sweeney predicted not only that within five to seven years, we would have not just PC-grade VR devices but also that these devices would have shrunk down into Oakley-style sunglasses.

    It will be interesting to see how the release of Apple’s first mixed reality headset, rumored for later this year, will move the needle on this.

  • Nico Grant, reporting for The New York Times:

    Last month, Larry Page and Sergey Brin, Google’s founders, held several meetings with company executives. The topic: a rival’s new chatbot… [ChatGPT] has shaken Google out of its routine… Google now intends to unveil more than 20 new products and demonstrate a version of its search engine with chatbot features this year, according to a slide presentation reviewed by The New York Times

    […]

    [Page and Brin] reviewed plans for products that were expected to debut at Google’s company conference in May, including Image Generation Studio, which creates and edits images, and a third version of A.I. Test Kitchen, an experimental app for testing product prototypes.

    […]

    Other image and video projects in the works included a feature called Shopping Try-on, a YouTube green screen feature to create backgrounds; a wallpaper maker for the Pixel smartphone; an application called Maya that visualizes three-dimensional shoes; and a tool that could summarize videos by generating a new one, according to the slides.

    […]

    Google executives hope to reassert their company’s status as a pioneer of A.I.

    While many of the rumored products don’t sound particularly compelling to me, Google does indeed seem serious about this bet. Although they recently laid off more than 12,000 employees, almost none of those employees were working in their AI division.

    I have no doubt that Google has all of the talent and resources required to become a leader in this space. The mystery is why they have been moving so slowly. Whether that is because of safety concerns, unclear monetization, or something else entirely is a question Google will need to sort out.

  • LangChain is an open source project designed to provide interoperability between large language models and external programs.

    From the project’s documentation:

    Large language models (LLMs) are emerging as a transformative technology… But using these LLMs in isolation is often not enough to create a truly powerful app - the real power comes when you are able to combine them with other sources of computation or knowledge… [LangChain] is aimed at assisting in the development of those types of applications.

    This looks like a super interesting project. I’ve talked before about how great it would be to combine ChatGPT with Wolfram Alpha. Well, that seems to be possible with LangChain. This Google Collab notebook and this HuggingFace project both appear to be examples of just that.

  • James Vincent at The Verge writes:

    In a glossy new video, [Boston Dynamics] has shown off its prototype Atlas robot tossing planks and tool bags around in a fake construction site.

    […]

    “We’re not just thinking about how to make the robot move dynamically through its environment, like we did in Parkour and Dance,” said Kuindersma. “Now, we’re starting to put Atlas to work and think about how the robot should be able to perceive and manipulate objects in its environment.”

    […]

    It’s a notable change in messaging from the Hyundai-owned company, which has never previously emphasized how its bipedal machines could be used in the workplace.

    In an announcement on the Boston Dynamics blog Calvin Hennick writes:

    While some Boston Dynamics robots, such as Spot and Stretch, are commercially available, Atlas is purely a research platform. The Atlas team focuses on pushing the forefront of what’s possible. The leaps and bounds forward in Atlas’ R&D can help improve the hardware and software of these other robots, while also advancing toward a “go anywhere, do anything” robot—capable of performing essentially all the same physical tasks as a person.

    James Vincent again:

    As ever, when parsing marketing materials from companies like Boston Dynamics, it’s important to notice what the company doesn’t say, as well as what it does. In this case, Boston Dynamics hasn’t announced a new product, it’s not saying it’s going to start selling Atlas, and it’s not making predictions about when its bipedal robots might work in factories. For now, we’re just getting something fun to watch. But that’s how Spot started, too.

    Do watch their YouTube video. It is, as with pretty much all of Boston Dynamics demonstrations, both super impressive and a little frightening. However, if you ever find yourself getting concerned about an imminent robot uprising you might find some solace in the fact that the robots would surely go after Boston Dynamics employees first.

  • Riley Goodside and Spencer Papay write:

    Anthropic, an AI startup co-founded by former employees of OpenAI, has quietly begun testing a new, ChatGPT-like AI assistant named Claude.

    […]

    Anthropic’s research paper on Constitutional AI describes AnthropicLM v4-s3, a 52-billion-parameter, pre-trained model… Anthropic tells us that Claude is a new, larger model with architectural choices similar to those in the published research.

    For context, GPT-3 has 175 billion parameters.

    Claude can recall information across 8,000 tokens, more than any publicly known OpenAI model, though this ability was not reliable in our tests.

    This is, effectively, how much “short-term memory” an AI model has. You definitely don’t want any information to be pushed out of memory during a normal chat session. Ideally, an AI model would remember information across multiple chat sessions although neither GPT-3 nor Claude have this ability at this time.

    Later in the article, the authors preform some comparisons between Claude and ChatGPT (GPT-3.5). Here are the big takeaways:

    • Both models are bad at math but Claude, at least occasionally, recognizes this fact and refuses to answer math problems when asked.
    • ChatGPT is quite good at code generation. The code Claude generates contains significantly more errors.
    • Both models appear to be broadly equivalent at logical reasoning tasks.
    • Both models are good at text summarization.

    The article concludes:

    Overall, Claude is a serious competitor to ChatGPT, with improvements in many areas. While conceived as a demonstration of “constitutional” principles, Claude feels not only safer but more fun than ChatGPT.

    And this is all with a model with somewhere around one third of the parameters of GPT-3? I have a feeling this is going to be an exciting year for LLM developments.

  • Kalley Huang, writing for The New York Times:

    It is now not enough for an essay to have just a thesis, introduction, supporting paragraphs and a conclusion.

    “We need to up our game,” Mr. Aldama said. “The imagination, creativity and innovation of analysis that we usually deem an A paper needs to be trickling down into the B-range papers.”

    […]

    Other universities are trying to draw boundaries for A.I. Washington University in St. Louis and the University of Vermont in Burlington are drafting revisions to their academic integrity policies so their plagiarism definitions include generative A.I.

    Maybe a future for essay writing looks more like:

    1. Craft an effective prompt for a given assignment.
    2. Read and fact check the initial output. Revise your prompt and return to step one as necessary.
    3. Taking into account things learned during the fact-checking process, revise and rewrite the output from step two. Cite external sources to support your claims.
    4. If your essay still fails an “AI detector” screening that means you have not revised it enough. Return to step three. If your essay contains factual inaccuracies or uncited claims, return to step three.

    Yes, this still assumes there will be reliable “AI detector” services. Yes, there will still be a cat and mouse game where students find ways to trick the AI detection systems. I don’t think that is really something you can avoid. So, sure, update your academic integrity policy accordingly. Ultimately, though, I think you need to start from the assumption that generative AI will be an ongoing presence in the classroom. From there, encourage a classroom culture that embraces AI as an imperfect, but increasingly important, tool.


    Previously:

  • Cedric Chin writes about the development of the original iPhone’s keyboard:

    Nobody on the 15-engineer team quite knew what the ideal software keyboard would look like. Over the next few weeks, the engineers developed a wide variety of prototypes. One developed a Morse-code-inspired keyboard which would have the user combine taps and slides to mimic dots and dashes. Another developed a piano-like keyboard where users would need to click multiple keys at once (hence the name) to type a specific letter. The remaining prototypes downsized the usual QWERTY keyboard, but these came with their own set of problems. The buttons were too small and there was no tactile feedback to tell the user whether they had hit or missed the button.

    This is a great illustration of how the most obvious solution, in hindsight, is often not at all clear in the moment.

  • Eleven Labs recently shared a demo of their new voice synthesis AI. It is worth listening to the audio samples. While I don’t think they are significantly better than the recent demo released by Apple, it is for precisely that reason that I think this is so noteworthy — the fact that small companies are able to build comparable offerings to the industry’s largest players is impressive.

    Also, I have to admit, their Steve Jobs voice simulation demo is impressive.

    Finally, as time goes on I am increasingly unable to understand why none of these recent advancements have trickled down into voice assistants. Why not hook up a speech recognition AI to GPT and then speak the result using one of these voice generation AIs? It must be inference cost, right? Otherwise, I must be missing something.

    Microsoft and OpenAI together could use Whisper, to ChatGPT, to VALL-E and dub it Cortana 2.0. Or put it in a smart speaker and instantly blow Amazon Alexa, Apple Homepod, Google Home offerings out of the water. And that is just using projects OpenAI and Microsoft released publicly!

  • I wrote in December about how ChatGPT could be improved by routing relevant questions to Wolfram Alpha — i.e. neuro-symbolic AI. It sounds like Stephen Wolfram has similar thoughts:

    There’ll be plenty of cases where “raw ChatGPT” can help with people’s writing, make suggestions, or generate text that’s useful for various kinds of documents or interactions. But when it comes to setting up things that have to be perfect, machine learning just isn’t the way to do it—much as humans aren’t either.

    […]

    ChatGPT does great at the “human-like parts”, where there isn’t a precise “right answer”. But when it’s “put on the spot” for something precise, it often falls down. But the whole point here is that there’s a great way to solve this problem—by connecting ChatGPT to Wolfram|Alpha and all its computational knowledge “superpowers”. 

    […]

    Inside Wolfram|Alpha, everything is being turned into computational language, and into precise Wolfram Language code, that at some level has to be “perfect” to be reliably useful. But the crucial point is that ChatGPT doesn’t have to generate this. It can produce its usual natural language, and then Wolfram|Alpha can use its natural language understanding capabilities to translate that natural language into precise Wolfram Language.

    These are exactly the types of informal integrations I expect to see in spades once we finally get a viable open source alternative to GPT.

  • Semafor:

    Microsoft has been in talks to invest $10 billion into the owner of ChatGPT… The funding, which would also include other venture firms, would value OpenAI… at $29 billion, including the new investment

    Gary Marcus:

    Whether you think $29 billion is a sensible valuation for OpenAI depends a lot of what you think of their future… On [one] hand, being valued at $29 billion dollars is really a lot for an AI company, historically speaking, on the other Altman often publicly hints that the company is close to AGI

    How much would AGI actually be worth? A few years back, PwC estimated that the overall AI market might be worth over $15 Trillion/year by the year 2030; McKinsey published a similar study, coming at at $13 trillion/year… If you really were close to being first to AGI, wouldn’t you want to stick around and take a big slice of that, with as much control as possible? My best guess? Altman doesn’t really know how to make OpenAI into the juggernaut that everybody else seems to think he’s got.

    Finally, Marcus shares some interesting information he received from an anonymous source:

    Turns out Semafor was wrong about the deal terms. If things get really really good OpenAI gets back control; I am told by a source who has seen the documents “Once $92 billion in profit plus $13 billion in initial investment are repaid [to Microsoft] and once the other venture investors earn $150 billion, all of the equity reverts back to OpenAI.” In that light, Altman’s play seems more like a hedge than a firesale; some cash now, a lot later if they are hugely successful.

    It is important to remember that OpenAI isn’t exactly a for-profit company but, instead, a “capped profit” company. From their press release announcing the new corporate structure:

    The fundamental idea of OpenAI LP is that investors and employees can get a capped return if we succeed at our mission… But any returns beyond that amount… are owned by the original OpenAI Nonprofit entity.

    OpenAI LP’s primary fiduciary obligation is to advance the aims of the OpenAI Charter, and the company is controlled by OpenAI Nonprofit’s board. All investors and employees sign agreements that OpenAI LP’s obligation to the Charter always comes first, even at the expense of some or all of their financial stake.

    Although at the end of the day OpenAI can always change course. From The Information:

    OpenAI has proposed a key concession as part of discussions with potential new investors. Instead of putting a hard cap on the profit sharing—essentially their return on investment—it could increase the cap 20% per year starting around 2025, said a person briefed on the change. Investors say this compromise, if it goes through, would make the deal more attractive because it would allow shareholders to obtain venture-level returns if the company becomes a moneymaker.

  • LAION-AI, the non-profit organization that created the original dataset behind Stable Diffusion, launched Open Assistant last week. From the project’s GitHub page:

    Open Assistant is a project meant to give everyone access to a great chat based large language model… In the same way that stable-diffusion helped the world make art and images in new ways we hope Open Assistant can help improve the world by improving language itself.

    Remember, LAION is not the company behind Stable Diffusion (that would be Stability AI), they just produced the training dataset. We have yet to see if they can build a successful product. They have genuinely exciting plans though!

    We are not going to stop at replicating ChatGPT. We want to build the assistant of the future, able to not only write email and cover letters, but do meaningful work, use APIs, dynamically research information, and much more, with the ability to be personalized and extended by anyone. And we want to do this in a way that is open and accessible, which means we must not only build a great assistant, but also make it small and efficient enough to run on consumer hardware.

    Whether or not LAION is able to accomplish their goals, I am optimistic that we will see serious developments in the open source large language model space this year.

  • John Naughton, writing for The Guardian:

    [ChatGPT] reminds me, oddly enough, of spreadsheet software, which struck the business world like a thunderbolt in 1979 when Dan Bricklin and Bob Frankston wrote VisiCalc, the first spreadsheet program, for the Apple II computer

    Eventually, Microsoft wrote its own version and called it Excel, which now runs on every machine in every office in the developed world. It went from being an intriguing but useful augmentation of human capabilities to being a mundane accessory

    Digital spreadsheets are perhaps the best example of a computational tool successfully augmenting the day-to-day work of a huge number of people. Spreadsheets have gone from nonexistent to simultaneously indispensable and mundane unbelievably quickly. If a similar augmentation occurs for prose it will be an equally, if not more, transformative development.

subscribe via RSS