2084: The Future of Subscriber Email: Introducing Newsreader: AI-Powered Breaking News Email Generation
I might have gotten a bit distracted and wrote an AI powered email generator which provides a daily summary of the latest and greatest developments.
So I was playing around with GPT-3’s summarize feature today. It’s a really cool feature where if you feed it a long article, it can condense it down into a few short sentences which make it easier to understand. And I thought that what if you ran it in two steps, one step to summarize the article, and the second to condense it down into a single email. And that’s exactly what I did to create Newsreader.
Newsreader scrapes a website - currently only The Guardian, TechCrunch and PapersWithCode, summarizes the latest news and sends you an email update. It’s really quick and scarily good. I modified the prompt so that the emails are upbeat and the results are really good in my limited and admittedly biased opinion.
The following is a transcript of one of the emails it generates:
Subject: Breaking News: AI, Unions, E-Cargo Bikes, Real Estate, Google, Amazon, SBM Bank India, TuSimple, and Cubzh Make Headlines!
Dear Reader,
Welcome to this week’s edition of ‘Tech News Roundup’. This week, we take a look at the latest developments in the world of tech, from AI-generated avatars to autonomous trucks and real estate emissions tracking.
First up, Lensa AI, the app making ‘magic avatars,’ has raised red flags for artists. Lensa acts as a middleman, charging users for the convenience of having the app do the heavy lifting. This has sparked a debate about the ethical implications of free, readily available AI tools and how they are poised to change entire industries. Artists have voiced their concerns about their art being lost in the sea of algorithmic copies, with some fearing that their experiences and the human process they represent could be worth less in the future.
Next, Microsoft could get its first official union as ZeniMax QA testers organize. QA testers at ZeniMax Media, which is owned by Microsoft, are attempting to form the first union at the company through the Communication Workers Alliance. QA workers are often expected to work unsustainable hours, known as “crunch”, and are paid lower than other game developers. Microsoft has promised not to stand in the way of union organizing, and the QA testers have opened a voting portal for their cause. If successful, it will be the largest U.S. video game union to date.
Onomotion has also raised €21 million to expand its e-cargo bike urban logistics business. The company aims to expand to several thousands of vehicles across Europe and North America, and also offer a fleet management program, over-the-air software updates and a logistics-as-a-service business model. 15-20% of their turnover this year will come from this service, which they plan to expand to German cities outside Berlin and Hamburg.
Accacia is tackling the real estate industry’s massive carbon emissions problem. The startup has raised $2.5 million in seed funding to expand across Southeast Asia, the Middle East, the United States and Canada. The funding was led by Accel and B Capital and included Blume Ventures, Good Capital, Zerodha’s Rainmatter Fund, Loyal VC and angel investors. Accacia's platform tracks carbon emissions from all investment asset classes and has been used to reduce direct emissions by 20% within the first six months of use.
Google has also introduced “Continuous Scrolling” on desktop for Search. This new feature can also provide more visibility to sites that did not rank highly enough to be on the first page. It comes as response to user feedback that Google search results are not always accurate. Google has also rolled out other features including surfacing results from Reddit and Quora, and testing a feature on the home screen to give users access to information like weather and stocks.
Amazon is set to launch Prime Gaming in India. This service offers free access to a number of titles and is bundled with Amazon Prime and Video plans. It also offers users in-game loot at no additional cost and a range of Twitch-focused features. Amazon had quietly removed some of the references to the launch in India late Monday, and did not comment on the request for comment.
SBM Bank India is also building its BaaS platform and seeks funding at $200 million valuation. The bank has been actively working with fintech startups, offering co-branded cards and powering neobanks. HDFC and ICICI Bank shares have significantly grown in value since 2010, and investors have shown an appetite for backing banks like Shivalik Small Finance Bank. SBM India's success may depend on retaining customers and deepening their partnerships.
TuSimple and Navistar have ended their partnership to co-develop self-driving trucks. The move comes after the return of TuSimple CEO Cheng Lu and amid concurrent probes and questions regarding a potential threat to U.S. national security. TuSimple has received close to 7,000 reservations for its Navistar trucks from customers, some of which will get refunds, while TuSimple is looking to bring on another OEM partner. The companies have not ruled out the possibility of working together in the future.
Finally, Cubzh wants to build the next-generation Minecraft. It has raised a $3 million seed round and is available as an alpha test. It is a C/C++ engine which can run on PC, macOS, iOS, Android and web browsers. Players can create objects, draw swords and vehicles, and those who can code can incorporate them in their games. The company may introduce web3 mechanisms for monetization, such as a utility token and object sales.
That’s all for this week’s edition of ‘Tech News Roundup’. We hope you found this week’s news interesting
Articles
Lensa AI, the app making ‘magic avatars,’ raises red flags for artists
Microsoft could get its first official union as ZeniMax QA testers organize
Onomotion raises €21 million to expand e-cargo bike urban logistics business
Accacia tackles the real estate industry’s massive carbon emissions problem
Google introduces “Continuous Scrolling” on desktop for Search
Amazon set to launch Prime Gaming in India
SBM Bank India, building BaaS platform, seeks funding at $200 million valuation
TuSimple and Navistar end deal to co-develop autonomous trucks
Cubzh wants to build the next-generation Minecraft
It’s a pretty good summary of what’s happening on the mainpage of TechCrunch. Now of course you can fiddle with the word count, or so forth, but I feel like this is a great way to collect all the various sources that you get your news from. Imagine if you could just select news sources from a list, and get a nice personalized email every day from Newsreader of all the news you find interesting on that day, and you could choose the length, the tone, the brevity, everything, It’ll be like having your own personal newsbot!
Beyond news, it can also summarize scientific articles. There are so many scientific articles out there today that it’s hard to keep track. In addition, abstracts are complex and hard to read, and the fundamental idea can get lost. But the AI can give you a summary of what’s important, so that you can understand what’s going on. It’s fascinating to see how far GPT-3’s knowledge and ability can stretch - a lot of the terms in these articles are newly defined terms, yet the produced output is entirely understandable. It is absurdly generalizable. The following is an example of the output it produced when I fed it the articles on paperswithcode.com
Unlock the Latest Advances in AI: Discover DINER, Coder-Reviewer Reranking, and NPM Now!
Welcome to the latest update on all the exciting developments in Artificial Intelligence research.
First, we have a new method for object compositing called Generative Models. This self-supervised framework can transform the viewpoint, geometry, color, and shadow of the generated object with no manual labeling. To improve the realism and faithfulness of the synthesized result images, the framework uses a content adapter to maintain categorical semantics and object appearance, as well as a data augmentation method.
Second, we have a vision for research and development in the field of Artificial Intelligence for the next decade and beyond. This vision is based on Active Inference, a formulation of adaptive behavior that can be read as a physics of intelligence which corresponds to maximizing Bayesian model evidence.
Third, a new method called DINER has been introduced for Depth-aware Image-based Neural Radiance fields. This method incorporates depth information into feature fusion and efficient scene sampling to enable larger viewpoint changes and higher synthesis quality.
Fourth, a new method called Coder-Reviewer reranking has been proposed to take code language models and rerank them with model likelihood. This method can improve the accuracy of the reranking with up to 17% absolute gain and works well with off-the-shelf hyperparameters when combined with executability filtering.
Fifth, a new framework has been developed to measure the distance from perfect calibration for probabilistic predictors. This framework suggests the development of three calibration measures – smooth calibration, interval calibration, and Laplace kernel calibration – which give quadratic approximations to the ground truth distance.
Sixth, the latest version of the ClueWeb datasets has been released, called ClueWeb22. It is far larger and more varied than earlier versions, making it suitable for commercial web search. Additionally, it provides various signals not available before, such as the visual representation of pages and pre-processed cleaned text.
Seventh, a new method called Coder-Reviewer reranking has been proposed to take code language models and rerank them with model likelihood. This method can improve the accuracy of the reranking with up to 17% absolute gain and works well with off-the-shelf hyperparameters when combined with executability filtering.
Eighth, a new method of training Vision Transformers (ViT) directly from the encoded features of JPEGs has been developed to avoid the decoding overhead of RGB images. With the improvements, the ViT-Ti model achieves faster training and inference with no accuracy loss compared to the RGB counterpart.
Finally, a new nonparametric masked language model called NPM has been introduced as an alternative to existing language models. It is trained with a contrastive objective and an in-batch approximation to full corpus retrieval and is shown to outperform significantly larger parametric models in zero-shot evaluation on a range of tasks.
We hope this update has been informative and has provided you with insights into the new developments in Artificial Intelligence. For more information, please follow the links provided.
Thank you for subscribing.
Sincerely,
The Newsreader
It’s pretty good, barring a few artifacts here and there, and as I said, pretty impressive given that these are brand new papers, and it’s a lot simpler than reading the abstracts involved. Without much more than a couple hours of work, this is already almost as good as most handwritten subscriber emails I get, and much quicker to generate. I think there’s quite a lot of potential in it, and especially for news site, it could be a good way to do aggregation.
This is still a work in progress, and if you want to get emails sent to you for your favorite news website, reach out, and I’ll send emails along and add the websites to the scraper. Try it out, see how it is. If there’s any glitches, do tell me.
This project though has made me realize the power of GPT-3. GPT-3 is the end of programming since it performs a lot of tasks by itself which programming is not capable of, or it takes a lot of time to do. I could probably handwrite a summary email generater, but GPT-3 does it much faster then I could, and better.
What this has made me realize is that the real way forward for programming isn’t necessarily Codex and the like. I think that tools which are essentially based on prompting and modifying GPT-3 and similar large models will become increasingly prevalent and will replace a lot of tasks which we use code for today. In the year 2084, code might be seen in the same way embedded system, or assembly language is today - something done by enthusiasts and specialists, with AI prompt engineering being the main way in which coding is done.