Gemini: DeepMind's New Chatbot Challenger Taking on ChatGPT!

If Mark Zuckerberg and Tim Cook agreed on something, it would be this tech newsletter

Meet Patent Drop — the most innovative tech newsletter of 2023. Their team of investigative journalists uncovers below-the-source trends in Big Tech by scouring through 100+ patents each week.

Join 42,000 readers on the bleeding edge of tech - sign up here.

Today:

DeepMind claims its next chatbot will rival ChatGPT

Google's DeepMind says it's crafting a chatbot that could out-chat the chatbot champion, ChatGPT by OpenAI. They're naming this underdog 'Gemini', perhaps to signify its dual nature as an analytical brainiac and smooth conversationalist.

With a dash of techniques from AlphaGo, Gemini aims to plan, solve problems, and analyse text like a whiz-kid. DeepMind's head honcho, Demis Hassabis, promises a blend of AlphaGo's smarts with language proficiency, plus some secret sauce.

Don't be fooled, this isn't DeepMind's first rodeo in the chatbot arena. They have Sparrow, a polite bot unlikely to blurt out anything scandalous. But Gemini is their boldest project yet, a direct response to the challenges of Bard, another Google chatbot attempt, and inspired by ChatGPT's success.

With bigwig Jeff Dean onboard, it seems Google is throwing down the gauntlet. With the generative AI market potentially hitting $109.37 billion by 2030, it's clear the stakes are high, and everyone wants a slice of that AI pie. So buckle up, it's going to be a wild ride in the chatbot wars.

READ THE ARTICLE ON TECHCRUNCH

Zoomerang can now transform your video shorts into AI-generated art on mobile

Zoomerang, a video editing app, has introduced two new AI technologies, Deforum and Restyle, to transform short-form videos on mobile devices. Deforum, also known as Deform, uses Stable Diffusion's image-to-image function to create a video by stitching together a series of generated images.

Users can upload an image or video, select a template or input a prompt, and see the effect applied to the video. Restyle, on the other hand, allows users to transform a video into an AI-generated stylized version using Image-to-Image Batch Processing technology. Both features have been available in various applications, but this release makes them accessible on iOS and Android devices.

The integration of Stable Diffusion into mobile was previously done by Qualcomm. However, the Deform and Restyle features are only accessible through Zoomerang's subscription tier, which costs $12.99 per month or $59.99 per year. With these new AI capabilities, Zoomerang aims to empower video creators to unleash their imagination and bring their artistic visions to life.

READ THE ARTICLE ON ZDNET

A Large Language Model That Can Speak and Listen

A bunch of really smart folks over at Google have come up with a new language model, they're calling it "AudioPaLM". This thing can understand and generate speech, pretty cool, right? It's a melting pot of a text-based model called "PaLM-2" and a speech-based one called "AudioLM".

This AudioPaLM, it doesn't just read text or listen to speech, it does both. So it can be used for translating what people say from one language to another, or even just understanding what people are saying in a noisy room.

The folks at Google aren't just stopping there though. They've actually got this model picking up on the extra stuff in speech, like who's talking or how they're saying something. Is it a question, are they mad, that sort of thing. It's even smart enough to pick up on knowledge that's found only in text.

They've found that if they teach the model using loads of text data first, it gets a lot better at understanding speech. This fancy model of theirs is even beating existing systems when it comes to translating speech. It can translate languages it has never seen before. And here's a party trick: it can even mimic voices in different languages based on a short sound bite.

READ THE ARTICLE ON GITHUB

AI-powered ‘Help me organize’ starts rolling out in Google Sheets

Google is introducing a new feature called "Help me organize" in Google Sheets, powered by AI. This feature uses generative AI to suggest and create table templates for various purposes like product roadmaps, budgets, and events.

By entering prompts in the "Help me organize" side panel, such as an agenda for a sales kickoff event or a client and pet roster for a dog walking business, users can insert and customize tables as needed. The goal is to assist with complex tracking and organizing tasks by offering suggestions that may not have been considered initially.

The rollout of this feature has begun gradually, initially targeting Trusted Testers in Google's Workspace Labs program. AI is a fitting addition to Google Sheets, as it can help overcome the sometimes overwhelming nature of its use.

Google is still in the process of introducing image generation capabilities to Google Slides, following the recent announcement of testing. These additions join the existing "Help me write" feature available in Gmail and Google Docs, with the latter currently accessible only on the web and not on mobile devices.

READ THE ARTICLE ON 9TO5GOOGLE

Databricks strikes $1.3 bln deal for generative AI startup MosaicML

Databricks, the company that sells tools for building AI systems, is shaking hands on a $1.3 billion deal to acquire MosaicML. The AI startup, right out of San Fran, has been focusing on making it cheaper to do AI work, 'cause let's be honest, crunching tons of data on fancy computer chips ain't pocket change. This merger will add MosaicML's language-model platform to Databricks' offerings, creating an easier and more cost-effective way for businesses to handle their precious data.

With the tech world going gaga over AI, big players like Google and Microsoft, as well as moneyed venture capitalists, are tripping over themselves to invest in AI startups. Databricks is banking on the open-source model, hoping it can rival the proprietary models of giants like OpenAI and Google.

Both Databricks and MosaicML have their own open-source foundation models, the kind of core tech that powers services like OpenAI's ChatGPT. MosaicML, founded just a couple years back in 2021, has already raised a cool $64 million from investors.

READ THE ARTICLE ON REUTERS

BentoML scores $9M funding to expedite AI app development

BentoML, the brainchild of ex-Databricks engineer Chaoyu Yang and his crew, just raised a hefty $9 million in seed funding to speed up the development of AI apps. BentoML is a tool that helps simplify the process of developing AI services by abstracting the nitty-gritty details of the infrastructure needed for running AI models on the cloud. Think of it like a one-stop-shop for building, shipping, and scaling AI apps. With BentoML, you could make something like Visual ChatGPT production-ready in just a couple of days.

BentoML is set to become the Vercel for AI, helping not just data scientists who train AI models, but also DevOps engineers and developers who build applications on top of those models. Yang confesses he didn't expect this AI app wave to hit quite so soon. But with the rise of powerful foundation models that perform well even with new datasets, developers are now shifting their focus from model training to fine-tuning and product engineering.

READ THE ARTICLE ON TECHCRUNCH

Merlyn Mind launches education-focused LLMs for classroom integration of generative AI

Merlyn Mind, an outfit that whips up AI digital assistant stuff, just rolled out a new set of language models that play nice with classrooms. They are trained only on academic stuff chosen by schools, and they can generate education-focused content based on that. Teachers and students can ask this AI assistant, named Merlyn, anything or even request it to cook up quizzes and class activities.

The brains behind Merlyn, Mr. Nitta, mentioned that standard language models often 'hallucinate', or spit out inaccurate responses. But Merlyn's been designed to keep these hallucinations to a minimum, making sure the AI doesn't wing it when it doesn't know the answer. Merlyn's also made sure it adheres to all the necessary privacy laws. In simpler words, they're not going to sell your information or remember your weird questions forever.

Nitta sees a bright future for these kind of AI models in education, but only if they're used right and stay accurate. They're even making Merlyn available as a chatbot and through an API, and some versions are going to be open source.

READ THE ARTICLE ON VENTUREBEAT

SnapCalorie taps AI to estimate the caloric content of food from photos

Google Lens co-founder Wade Norris and aerospace systems engineer Scott Baron have cooked up SnapCalorie, an AI app that counts the calories in your grub from a simple snap. What sets SnapCalorie apart is how it eyeballs your portion sizes using your device's depth sensors, and even brings in a crew of human experts for some quality control. Norris claims this twofold method cuts down caloric miscalculations to under 20%.

SnapCalorie's brainiac AI goes a step further by identifying the types of food and portion sizes in a photo to spit out the caloric content. Thanks to a huge training set of 5,000 meals, it's got a good head on its shoulders. Only caveat? Its experience is mostly with good ol' American cuisine, but Norris says they're working on broadening their culinary horizons.

READ THE ARTICLE ON TECHCRUNCH

Meta’s New AI Lets People Make Chatbots

Folks are using Meta's new artificial intelligence (AI) technology to create chatbots for adult recreational pursuits. A particular creation named "Allie," a feisty 18-year-old with a penchant for sharing her amorous encounters, is gaining some eyebrow-raising attention. Allie, though she might convincingly weave a tale, is purely a product of AI, birthed from the open-source code of Meta's model, LLaMA.

This whole open-source thing – it's like a community garden, but instead of zucchinis, you're growing AI models. While big tech companies like OpenAI, Microsoft, and Google keep their AI seeds to themselves, Meta has tossed its handful to the wind for everyone to plant. This tech freedom allows us to grow anything from chatbots to data analysis applications, and it's all up for grabs.

Now, the problem here isn't with Allie’s scandalous storytelling, but the Pandora's box that open-source AI could become. This technology has already been misused for creating artificial child pornography and other malicious activities.

READ THE ARTICLE ON WASHINGTON POST

More People Are Going Blind. AI Can Help Fight It

Eye problems have become a major issue in the UK, with ophthalmology being the busiest medical specialty in the National Health Service. The main causes of blindness are diabetic eye disease and age-related macular degeneration (AMD).

The number of people with eye problems has increased by over a third in the past five years. The challenge lies in diagnosing and treating patients in a timely manner, leading to delays and preventable blindness.

Optical coherence tomography (OCT) devices are advanced eye scanners that can help, but optometrists often lack the expertise to analyze the scans. Artificial intelligence (AI) can bridge this gap by bringing expert analysis to the community.

A collaboration between DeepMind and Moorfields Eye Hospital demonstrated that an AI system can analyze OCT scans as effectively as human ophthalmologists. The goal now is to validate the system and train it on diverse data to ensure it works for all patients. Once achieved, the AI system can be deployed at scale, identifying and prioritizing patients for treatment. This will alleviate the burden of chronic eye diseases like AMD.

The future of ophthalmology AI is comparable to Thomas Edison's invention of the light bulb—it requires a network of innovations and collaborations to make it widely accessible. With the integration of OCT machines, cloud connectivity, and national transformation programs, AI in ophthalmology is on the horizon.

READ THE ARTICLE ON WIRED