The AI Daily Brief: Artificial Intelligence News and Analysis - Anthropic Has (Maybe) Solved a Holy Grail of Business AI

Episode Date: November 28, 2024

Anthropic’s latest update introduces customizable writing styles for Claude AI, offering the potential to mimic individual or brand-specific tones. This innovation could unlock new business use cases, especially for creating content at scale. But how effective is it? Brought to you by: Vanta - Simplify compliance - ⁠⁠⁠⁠⁠⁠⁠https://vanta.com/nlw Plumb - AI automation that just works - https://useplumb.com/ The AI Daily Brief helps you understand the most important news and discussions in AI. Subscribe to the podcast version of The AI Daily Brief wherever you listen: https://pod.link/1680633614 Subscribe to the newsletter: https://aidailybrief.beehiiv.com/ Join our Discord: https://bit.ly/aibreakdown

Today on the AI Daily Brief, a new anthropic feature that promises to be able to copy your own writing style. Before that in the headlines, how AI is helping fix one of the most painful parts of Thanksgiving. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. To join the conversation, follow the Discord link in our show notes. Welcome back to the AI Daily Brief Headlines edition, all the daily AI news you need in around five minutes. We are coming right up on Thanksgiving. In fact, by the time you're listening to this, it may be Thanksgiving. And it turns out that there is a very specific experience that Americans really, really hate on this day.

It's not getting stuck in conversation with a weird drunk relative, but has to do with the cooking process. For over 50 years, Butterball has maintained a turkey hotline. Think 9-1-1 for turkey-related emergencies. This very famously featured in an episode of the West Wing, where President Josiah Bartlett mentioned offhandedly that there should be be a special service for helping people with their Thanksgiving turkeys, only to discover that one existed. He then, of course, tried to economists explain to them how to make a turkey, but that's neither

here nor there. The point is that Butterball has decades of audio data from customers explaining the snags they run into on Thanksgiving Day. By piping this data through an AI summarization tool, Butterball came up with a startling finding. People hate dealing with frozen turkey. Every year, Thanksgiving chefs call up in a state of panic when they realize they've forgotten to thaw the bird. While this isn't a surprise, the key insight was that complaints about thawing have exploded in recent years, something that Butterball might not have noticed without the help of AI to summarize millions of customer service interactions. The firm has now undertaken a three-year process of revising their data structure and analyzing customer input and is now quite possibly the most AI-forward turkey processor in the world.

This is all they do, and each Thanksgiving day has 18 months of preparation put into it from procurement to logistics to customer service. AI is also helping with those tasks allowing the firm to better optimize their internet. inventory and deliveries. The crown jewel of their AI overhaul, however, is the thawless turkey, making its debut this Thanksgiving. That product involved hours and hours of testing with turkey scientists collecting hundreds of data points. With the help of machine learning, Butterball believes they have engineered the perfect thawless turkey. They've paired it with precise cooking instructions to ensure it comes out right every time. So, if you are cooking a thawless turkey this year, you could be eating your first AI-enhanced Thanksgiving meal.

Next up, perplexity is about to fall into one of the single greatest traps that any startup can fall into, which is trying to make a cool next-generation hardware device. Earlier this week, CEO Aravon Shrinivas wrote, considering making a simple under-50-dollar hardware device that will reliably answer your questions voice-to-voice. Just do this, but do it very well. If this post gets more than 5,000 likes, we'll definitely make it. When the post did indeed get more than 5,000 likes, he followed up with the comment, all right, LFG. AI hardware, though, has been a fairly fraught category over the past year. The arguable most successful product, the Rabbit R1,

launched two very lukewarm reviews. The company claims to have shipped 130,000 units, but they are readily available at a steep discount in the secondary market. Others, like the Humane AI Pin, had awful reviews, weak sales, and a recall that has forced the company to look for an acquirer. Still, if the last year has taught us anything, it's don't fade perplexity. The company is reportedly in the middle of raising a half a billion dollars and has been shipping relentlessly for months.

AI hardware is also a theme that could rise in popularity next year. Mid Journey recently formed a hardware team. OpenAI is looking to develop a device designed by Johnny Ive, and Rabbit's agentic upgrade is showing some promise. Still, shipping an AI wearable at a $50 price point seems pretty tough. And some are trying to convince Trinnavasa do literally anything else. Solopreneur Peter Levels wrote, in my opinion, don't do it. It's already done and never works. We don't need another AI device. We already have a smartphone. Just double down on making the mobile app the best ever. Add a great competitor to Google Images because I use that a lot, but it should be a similar layout as LLM-style layout doesn't work for it. And I think this is the sentiment from lots

of people. Taking on Google and finally being the first company to actually have a chance to disrupt search in 20 years seems like a big enough task. Others, though, are egging him on, with Boyan Tungu saying, I already made one for $200 using off-the-shelf PC and chat GPT advanced mode. I'm about to make one with Raspberry Pi for around $100. So will it actually happen? We will just have to wait and see. Over in funding news, a group of former Google and Stripe executives have raised $56 million to build an operating system for AI agents. The startup called dev slash agents is led by a group of founders who helped build the Android platform.

They're now applying the same playbook to AI agents. The key insight is that agents will need a common technical framework to connect to services and communicate with each other, much like different apps within an operating system. Co-founder and CEO David Singleton said, We need an Android-like moment for AI. We can see the promise of AI agents, but as a developer, it's just too hard to build anything good. The goal is to create a new user interface that allows more natural interactions with agents

across different devices. One thing that's for sure is that the team is absolutely stacked with talent. Hugo Barr, the startup's chief product officer, and formerly VP of product management for Android said, this is a team that's built the last three generations of operating systems. Investors are certainly excited. Nina Acheji on a partner at Index Venture said, if you think about the people at this company and founder market fit, it couldn't be more relevant for what they've set out to go build.

Jill Chase, a partner at Capital G, said, this is a once-in-a-generation opportunity they're attacking. Announcing the company and laying out his vision, Singleton tweeted, Modern AI will fundamentally change how people use software in their daily lives. Agendic applications could, for the first time, enable computers to work with people in much the same way people work with people. But it won't happen without removing a ton of blockers. We need new UI patterns, a reimagined privacy model, and a developer platform that makes

it radically simpler to build useful agents. That's the challenge we're taking. on. We're building a cloud-based OS for trusted agents to work with users across all of their devices. We want to help people spend time on what matters to them. In another bit of funding, news, Black Forest Labs is in talks to raise $200 million in their first major funding round. The German startup is only a few months old, but has garnered a claim for their flux text to image model. The model is driving image generation for XAI's Grock Chatbot and is generally considered to be right at the top of state-of-the-art. The company's founding team includes

several computer scientists involved with creating stable diffusion. The new funding round is rumored to value the company at a billion dollars and will be led by A16Z. Their seed round was raised in August, gathering 31 million from A16Z, and including Oculus co-founder Brendan Aribay and Y Combinator head Gary Tan. In product news, Google is connecting Spotify to Gemini. According to code spotted in the Gemini extension, the AI assistant will soon be able to take control of Spotify. Users will be able to use AI to search and play music using natural language requests.

For the moment, Gemini won't be able to make playlists or interact with radio stations on the platform. This marks the second integration for Gemini outside of Google's apps with WhatsApp compatibility added last month. And the question is whether this reveals something of Google's upcoming strategy. Is their focus going to be on bringing an agenic type experience to a wide constellation of apps? It would certainly be in line with how they've done things in the past. Lastly, today, a new business line for Uber, the company is launching a new AI data labeling service. The new division called Scaled Solutions has begun hiring contract workers to complete data labeling tasks. The initiative builds on an internal team that tackles large-scale annotation tasks for the

ride-sharing company, but the division will now offer their service to external clients. Data labeling is an unglamorous but rapidly growing part of the AI industry. Scale AI, a startup that offers similar services, is currently valued at $14 billion among the top tier of venture-backed companies in the space. We're also seeing high-quality data labeling emerge as a superpower for some model builders. Last month, a new video model from Chinese lab Minimax blew the industry away with its unprecedented capabilities. And some suspected the secret to building that performant video model was plentiful and accurately labeled training data. Regarding plans for this new division, a NUBris spokesperson said, having performed these tasks at scale over the past decade as part of our own growth,

we deeply understand the needs of companies requiring these services. They added that hiring independent contractors aligns, quote, with our expertise as one of the world's largest providers of flexible work opportunities. It may not be big news from the frontier labs, but there is still plenty going on as we head into this holiday week? For now, that is going to do it for the AI Daily Brief Headlines edition. Next up, the main episode. Today's episode is brought to you by Plum. Want to use AI to automate your work but don't know where to start? Plum lets you create AI workflows by simply describing what you want. No coding or API keys required. Imagine typing out, AI, analyze my Zoom meetings and send me your insights in Notion and watching it come to life before your eyes.

Whether you're an operations leader, marketer, or even a non-technical founder, Plum gives you the power of AI without the technical hassle. Get instant access to top models like GPT40, Claude Sonnet 3.5, assembly AI, and many more. Don't let technology hold you back. Check out Use Plum, that's Plum with a B, for early access to the future of workflow automation. Today's episode is brought to you by Vanta.

Whether you're starting or scaling your company's security program, demonstrating top-notch security practices, and establishing trust is more important than ever. Venta automates compliance for ISO-27-01, SOC2, GDPR and leading AI frameworks like ISO 402,001, and NIST AI Risk Management framework, saving you time and money while helping you build customer trust. Plus, you can streamline security reviews by automating questionnaires and demonstrating your security posture with a customer-facing trust center, all powered by Vanta AI.

Over 8,000 global companies like Langchain, Lila AI, and factory AI use Vanta to demonstrate AI trust and prove security in real time. Learn more at Vanta.com slash NLW. That's vanta.com slash NLW. Today's episode is brought to you by Super Intelligent. Every single business workflow and function is being remade and reimagined with artificial intelligence. There is a huge challenge, however, of going from the potential of AI to actually capturing that value. And that gap is what Super Intelligence is dedicated to filling.

Super Intelligence accelerates AI adoption and engagement to help teams actually use AI to increase productivity and drive business value. An interactive AI use case registry gives your company full visibility, into how people are using artificial intelligence right now. Pair that with capabilities building content in the form of tutorials, learning paths, and a use case library. And super intelligent helps people inside your company show how they're getting value out of AI while providing resources for people to put that inspiration into action. The next three teams that sign up with 100 or more seats are going to get free embedded

consulting. That's a process by which our super intelligent team sits with your organization, figures out the specific use cases that matter most to you, and helps actually ensure support for adoption of those use cases to drive real value. Go to B-Super.a.I. to learn more about this AI enablement network. And now back to the show. Welcome back to the AI Daily Brief. There is a very common pattern when people start interacting with generative AI.

And basically that pattern goes something like, at first you are absolutely blown away by capabilities. Whether it's an image generator like mid-jurney, a voice synthesizer like 11 labs, or, of course, at LLM like ChatGPT or Claude, the things that AI can do make you feel like a wizard. However, inevitably, the deeper that you get, the more you find things that aren't quite right

or are limiting in terms of just how far you can take it. A great example in the image generation space is consistent characters. It's wonderful if you're just trying to create one-off images. You can get incredible fidelity and specificity in terms of what you want it to look and feel like. But if you're trying to do that across an entire range of images, if you're trying to create the basis for animation or a comic

book, it gets a lot harder. And lots and lots of effort has been spent both from the big image generation companies as well as third parties on how to do that better. And as that capability comes online, the key thing is that it unlocks an entire set of use cases that were otherwise cut off. In the LLM space, one of those holy grails, in other words, one of the types of updates that could unlock a huge number of new use cases or just improve what the LLM is already being used for in a fundamental and significant way is the ability for an LLM to imitate a particular writing style. One of the things that people figured out with ChatGPT and Claude quite quickly was that there is a particular flavor and feel to LLM generated writing. There are certain words that get used

way more by LLMs than get used by real people, words like delve, and just in general, there's a particular style that feels recognizable. Now, there are ways that you can use prompts to try to get around that. You can coach the LLM to write in a particular way. You can give it references. But one of the things that people have most wanted is the ability to just upload a set of their own documents and have the LLM be able to natively copy them. In fact, I've tried extensively to do this. I built a custom GBT, which as you can see has like 15 reference documents for from short essays to long-form writing, and the short of it is that it's only okay. It definitely makes ChatGPT not sound like ChatGPT,

and it certainly has some patterns that mimic how I write, but it also doesn't really sound like me, certainly not enough that I'm going to outsource important writing to it in any sort of short order. Other companies recently have tried to productize versions of this. Spiral, for example, is a product from Every, which does something in this space. their approach is really interesting. Basically, they're trying to make it easy for you to translate one type of content to another.

So I can say I'm generating YouTube videos, podcasts, and LinkedIn posts. And then it shows you how it can take that source work and turn it into lots of other posts. Now, part of the process is that it tries to retain style as it builds out this set of additional assets. So, for example, they let you play with historical examples as a way to demonstrate what it can do. And yet still, what people have really. really wanted is this to be built natively into the LLMs themselves and for it to be good. And so when Anthropic announced this new style feature yesterday, people in the know really got excited. Matt Wolfe, for example, wrote, I feel like this is a feature that everyone's been

wanting and waiting for. Unfortunately, this news has been overshadowed by the SORA leak, which has already been shut down. This is bigger, in my opinion. Anthropic, for their part, announces, with styles you can now customize how Claude responds. Select from the new preset options, concise, explanatory, or formal. And so there are actually two parts of this feature. The first is that there is just a built-in style selector where any prompt that you have, you can say if you want it to be normal, the default responses from Claude,

concise, shorter responses and more messages, explanatory, educational responses for learning, or formal, clear and well-structured responses. That on its own is a really great feature. It allows you to not have to prompt to be clear about the style that you want. It's instead now built into the UI. However, the real juice and what's got people excited is the fact that you can create and edit styles. So what you do is you select the styles menu, press create and edit styles, and then create custom style.

From here, you can add a writing example, which can be a document, or you can copy paste text, or you can simply describe the style you're going for. I added a recent post from LinkedIn that I had written that was a little bit longer and more substantial, more blog posty, and what it came back with was a style that it called Tech Translator. The style summary was deliver analytical insights through conversational and authoritative communication. So now when I go back, let's say write a short blog post

about AI agents in the enterprise. emphasize that while they might not be production ready yet, 20205 is likely the year where people start integrating agents at their companies and forward-thinking enterprises should get ahead. Let's copy this. Let's do it normal first. And you get this piece, AI agents in the enterprise, preparing for the 2025 wave.

This, because it's clawed, sounds nothing like me, it has nothing to do with how I would write. Just by way of one example, the paragraphs are much longer than would make sense for the type of style that I'm going for. Now let's try it with the tech translator style, the style that it created for me. This is certainly not perfect. It's a little more tweet than I'd like to think that I am. The first paragraph is, let's get real about AI agents in the enterprise. While everyone's watching demo videos and proof of concepts, most companies are still sitting on the sidelines. And honestly, that's been the right move, until now.

that is so much closer to how I write. Short, crisp sentences, punchy and trying to get people to pay attention. Now, it's very hard, of course, to describe in the context of a podcast whether this works better. But my point is that at least in my initial trials, this gets a heck of a lot closer natively than the stuff I've done before. And in the context especially of imagining a use case where a company is trying to mass produce content, this is such a huge unlock in terms of the average quality of a piece of writing that's going to come from Claude. No, not everyone has been super impressed. Matt Wolfe got a style summary that said,

deliver technical AI news with enthusiastic conversational expertise that makes complex topics engaging and accessible. But he said, unfortunately, I'm kind of disappointed. I uploaded 90 minutes worth of transcripts, and I don't think this reads like me at all. Luckily, I can give feedback and try to get it even closer. One thing I will point out is that Matt uploaded YouTube transcripts, which are him talking.

That's different than writing. And I wouldn't be surprised if that's a tougher translation for this technology. But regardless of what the source is, the point is that it's important to not overstate how close this hues to an individual style. I think the real value, and where a lot of people are going to get excited, is a little bit copying their own style, but a lot being able to create a bank of existing style presets that make it much faster to toggle and get the exact type of output you want

for any particular given use case. This to me perfectly encapsulates where we are with the development of LLMs. It is not just about pushing the state of the art. It is also about user experience and making these things actually work on a personal and in a business context. With that in mind, I think writing styles is a huge upgrade and one that I'm excited to play with more. For now that that's going to do it for today's AI Daily Brief. Until next time, peace.