Tag: machine learning

Posted on

Low-code ML platform Predibase raises another $12.2M

End-to-end machine learning platform Predibase today announced a $12.2 million expansion to its $16.25 million Series A funding round from last year. The company also announced that its low-code, declarative ML platform for developers is now generally available.

During the beta period, which launched when the company came out of stealth last year, users have trained over 250 models on the platform. Now that the service is generally available, these users can also use Predibase to deploy their own large language models (LLMs) instead of using an API from the likes of OpenAI. Users will also get access to Predibase’s own LudwigGPT LLM — named after the suite of machine learning tools Predibase co-founder Piero Molino launched in 2019 (and not the tragic 19th-century Bavarian king).

“Every enterprise wants to gain a competitive edge by embedding ML into their internal and customer-facing applications. Unfortunately, today’s ML tools are too complex for engineering teams, and data science resources are stretched too thin, leaving the developers working on these projects holding the bag,” said Piero Molino, co-founder and CEO of Predibase. “Our mission is to make it dead simple for novices and experts alike to build ML applications and get them into production with just a few lines of code. And now we’re extending those capabilities to support building and deploying custom LLMs.”

To do this, the company also today announced its Data Science Copilot, a system that can give developers recommendations on how to improve the performance of their models. Predibase is also launching a free two-week trial version of its platform.

Current customers include the likes of Paradigm (which does something with crypto) and Koble.ai (a tool that uses AI to help investors find early-stage investments).

Like most startups at this stage, Predibase plans to use the new funding to expand its go-to-market functions and build out its platform.

Between low-code/no-code ML platforms from the likes of AWS, Google and Microsoft and plenty of startups in this space, Predibase is operating in an increasingly crowded market. The company argues that its focus on developers and its ability to provide them with easy escape hatches from the low-code environment allows it to stand out.

Predibase exits stealth with a low-code platform for building AI models

Low-code ML platform Predibase raises another $12.2M by Frederic Lardinois originally published on TechCrunch

Posted on

SUNDAY REWIND: Robotics and society by Cynthia Yeung

This week’s Sunday Rewind is a #mtpcon SF+Americas keynote from Cynthia Yeung, Head of Product at Plus One Robotics. It’s a talk that gets us thinking about human/robot interaction and how we might discover a deeper understanding of our own humanity through robotics and AI. Cynthia reflects on the physical, economic and social context of Read more »

The post SUNDAY REWIND: Robotics and society by Cynthia Yeung appeared first on Mind the Product.

Posted on

Union AI raises $19.1M Series A to simplify AI and data workflows with Flyte

Union AI, a Bellevue, Washington–based open source startup that helps businesses build and orchestrate their AI and data workflows with the help of a cloud-native automation platform, today announced that it has raised a $19.1 million Series A round from NEA and Nava Ventures. The company also announced the general availability of its fully managed Union Cloud service.

At the core of Union is Flyte, an open source tool for building production-grade workflow automation platforms with a focus on data, machine learning and analytics stacks. The idea behind the platform was to build a single platform that teams can then use to create their ETL pipelines and analytics workflows, as well as their machine learning pipelines. And while there are other projects on the market that offer similar orchestration capabilities, the idea here is to build a tool that is specifically built for the needs of machine learning teams.

Flyte was originally developed inside of Lyft, where Union AI CEO and co-founder Ketan Umare developed some of the company’s earliest machine learning–based ETA and traffic models in 2016. At the time, Lyft had to glue together various open source systems to put these models into production.

“We got something running, but behind the scenes, it was a man behind the curtain. It was happening, but it was a lot of work,” Umare said. “What we learned was that other teams in the company were also struggling — and these were massive teams. And what happens when teams struggle is that they cannot keep the talent on. That’s a big problem, but what was the root of that? They were not able to deliver their things and they were not able to articulate why they were not able to deliver. It turns out to be an infrastructure problem.”

Image Credits: Union.ai

So he set out with a small team to build out the infrastructure tooling to make it easier for these teams to build their models and put them into production. But there was always friction between the software engineers and machine learning specialists. “The reason was that — at least in the way I have distilled it — I think software and machine learning systems or AI products are inherently different beasts,” Umare argued. In his view, software typically matures over time while AI models tend to deteriorate. These models, he noted, also often change based on external factors that users have little control over. “So you cannot use the same infrastructure that you use for [software deployments],” he said.

At that point, the team decided to open source its work in the form of Flyte and work with others to build out a more machine-learning-native platform.

As is so often the case, Umare and four other members of the original Flyte team then decided to build a startup around these core ideas and the Flyte open source project, with Union AI launching in late 2020.

Currently, Flyte is being used by companies like blackshark.ai, HBO, Intel, LinkedIn, Spotify, Stripe, Wolt, and ZipRecruiter.

“The fun thing about working with these large companies — what we do in the open source — is that we are working on some of the biggest models on our platform. So we know it works and we didn’t have to build anything specifically because we’ve been doing this for years. We just had to extend a couple of things,” Umare said.

“Based on a single team, we see 10x more offline training jobs dispatched from Flyte, and that results in 5x more frequent model releases with sizable business gains,” said Mick Jermsurawong, a machine learning infrastructure engineer with Stripe. “I think the realization here is that ML productivity is not a nice to have but actually a business requirement.”

But the Union AI platform isn’t simply building Flyte-as-a-service. The team also built Pandera (a framework for data testing) and Union ML (a framework that sits on top of Flyte and helps teams build and deploy their models using their existing set of tools). Union Cloud combines all these elements and layers a set of enterprise tools, such as single sign-on, on top of it.

“Machine learning, and especially large language models, raise big issues around privacy and information security. Companies are becoming increasingly wary of using services where they lose control over what precisely happens with their data,” said Greg Papadopoulos, venture partner, NEA. “Combining the power of big models with rich company data has to be handled with care — that’s one of the reasons why we’re so excited about the progress made by the Union.AI team, first with Flyte and now with Union Cloud. This is exactly what people are demanding and a real differentiator: Let me exploit the power of large language models while maintaining control and ownership of my data.”

Union AI raises $19.1M Series A to simplify AI and data workflows with Flyte by Frederic Lardinois originally published on TechCrunch

Posted on

Visual Layer helps enterprise manage the massive visual data sets they need to build AI models, raises $7M

Training machine learning models for computer vision use cases takes massive amounts of images. Often, those images are mislabeled, broken or duplicated, leading to sub-par model performance. But with millions of images in many datasets, it’s virtually impossible to catch these issues. Visual Layer, a Tel Aviv-based startup that wants to enable data scientists and ML engineers to find these issues before they impact their models, today announced that it has raised a $7 million seed funding round led by Madrona and Insight Partners.

The company built a system that, without relying on expensive GPUs, can analyze hundreds of millions of images and automatically find potential issues within these data sets. At the core of Visual Layer’s technology stack is the open-source fastdub project.  The company’s co-founders Danny Bickson (CEO) and Amir Alush (CTO) developed this project based on their experience at companies like Apple, where Bickson was previously a Str Data Science Manager of the company acquired the AI startup Turi (which he co-founded), and Brodmann17, which Alush co-founded. Fastdup, which recently hit version 1.0, helps engineers find potential issues in their image data sets, clusters and visualizes them.

Image Credits: Visual Layer

In their research, the team (which also includes former Turi CEO and co-founder Carlos Guestrin among the co-founders) found that the popular ImageNet-21K pre-training dataset includes over a million pairs of duplicates among its just over 14 million images. Most datasets also include broken images or very similar images that have vastly different labels. Often, thousands of images are also simply mislabeled and are then being used to train the wrong model (think images of shoes being used to train a face detection model).

“Companies and organizations across the globe are experiencing the explosion of data, and visual data is one of the most complex and challenging data types to manage. Understanding, curating and managing this content is crucially important to build meaningful services for customers in a broad set of industries – from retail to manufacturing to self-driving cars and more,” said Bickson. “Companies are struggling with those huge amounts of data, they often have no clue where their data is and what is inside it. They develop their homegrown tools since there is no infrastructure and no common standards.”

Unsurprisingly, Visual Layer is essentially offering fastdub as a service (with additional enterprise features layered on top).

Image Credits: Visual Layer

Some of the company’s early users include the likes of Indian social commerce platform Meesho, which used fastdup to improve the quality of its image gallery of 200 million products, John Deere, Honeywell, Winnow and Nuvilab.

“Despite the idea that bigger datasets mean better models, when it comes to images and video, messy underlying datasets can produce suboptimal models and error-prone results. Now with the reality of large-scale AI models we must solve the data problem. The immediate excitement we saw after the launch of fastdup made clear to us that customers agree. We are excited to work with the Visual Layer team and the fastdup community to build a new, foundational component of the AI application stack,” said Jon Turow, Partner at Madrona.

Visual Layer helps enterprise manage the massive visual data sets they need to build AI models, raises $7M by Frederic Lardinois originally published on TechCrunch

Posted on

Apple reveals new accessibility features, like custom text-to-speech voices

Apple previewed a suite of new features today to improve cognitive, vision and speech accessibility. These tools are slated to arrive on the iPhone, iPad and Mac later this year. An established leader in mainstream tech accessibility, Apple emphasizes that these tools are built with feedback from disabled communities.

Assistive Access, coming soon to iOS and iPadOS, is designed for people with cognitive disabilities. Assistive Access streamlines the interface of the iPhone and iPad, specifically focusing on making it easier to talk to loved ones, share photos and listen to music. The Phone and FaceTime apps are merged into one, for example.

The design also is made more digestible by incorporating large icons, increased contrast and clearer text labels to make the screen more simple. However, the user can customize these visual features to their liking, and those preferences carry across any app that is compatible with Assistive Access.

As part of the existing Magnifier tool, blind and low vision users can already use their phone to locate nearby doors, people or signs. Now Apple is introducing a feature called Point and Speak, which uses the device’s camera and LiDAR scanner to help visually disabled people interact with physical objects that have several text labels.

Image Credits: Apple

So, if a low vision user wanted to heat up food in the microwave, they could use Point and Speak to discern the difference between the “popcorn,” “pizza” and “power level” buttons — when the device identifies this text, it reads it out loud. Point and Speak will be available in English, French, Italian, German, Spanish, Portuguese, Chinese, Cantonese, Korean, Japanese and Ukrainian.

A particularly interesting feature from the bunch is Personal Voice, which creates an automated voice that sounds like you, rather than Siri. The tool is designed for people who may be at risk of losing their vocal speaking ability from conditions like ALS. To generate a Personal Voice, the user has to spend about fifteen minutes reading randomly chosen text prompts clearly into their microphone. Then, using machine learning, the audio is processed locally on your iPhone, iPad or Mac to create your Personal Voice. It sounds similar to what Acapela has been doing with its “my own voice” service, which works with other assistive devices.

Acapela lets anyone back up their own voice for free in minutes — just in case

It’s easy to see how a repository of unique, highly-trained text to speech models could be dangerous in the wrong hands. But according to Apple, this custom voice data is never shared with anyone, even Apple itself. In fact, Apple says doesn’t even connect your personal voice with your Apple ID, since some households might share a log-in. Instead, users must opt in if they want a Personal Voice they make on their Mac to be accessible on their iPhone, or vice versa.

At launch, Personal Voice will only be available for English speakers, and can only be created on devices with Apple silicon.

Whether you’re speaking as Siri or your AI voice twin, Apple is making it easier for non-verbal people to communicate. Live Speech, available across Apple devices, lets people type what they want to say so that it can be spoken aloud. The tool is available at the ready on the lock screen, but it can also be used in other apps, like FaceTime. Plus, if users find themselves often needing to repeat the same phrases — like a regular coffee order, for example — they can store preset phrases within Live Speech.

Apple’s existing speech-to-text tools are getting an upgrade, too. Now, Voice Control will incorporate phonetic text editing, which makes it easier for people who type with their voice to quickly correct errors. So, if you see your computer transcribe “great,” but you meant to say “grey,” it will be easier to make that correction. This feature, Phonetic Suggestions, will be available in English, Spanish, French and German for now.

Image Credits: Apple

These accessibility features are expected to roll out across various Apple products this year. As for its existing offerings, Apple is expanding access to SignTime to Germany, Italy, Spain and South Korea on Thursday. SignTime offers users on-demand sign language interpreters for Apple Store and Apple Support customers.

Apple adds live captions to iPhone and Mac, plus more accessibility upgrades to come

Apple puts accessibility features front and center

Apple reveals new accessibility features, like custom text-to-speech voices by Amanda Silberling originally published on TechCrunch

Posted on

Together raises $20M to build open source generative AI models

Generative AI — AI that can write essays, create artwork and music, and more — continues to attract outsize investor attention. According to one source, generative AI startups raised $1.7 billion in Q1 2023, with an additional $10.68 billion worth of deals announced in the quarter but not yet completed.

There’s scores of competition, including incumbents like OpenAI and Anthropic. But despite that fact, VCs aren’t shying away from untested players and up-and-comers.

Case in point, Together, a startup developing open source generative AI, today announced that it raised $20 million — on the larger side for a seed round — led by Lux Capital with participation from Factory, SV Angel, First Round Capital, Long Journey Ventures, Robot Ventures, Definition Capital, Susa Ventures, Cadenza Ventures and SCB 10x. Several high-profile angel investors were also involved, including Scott Banister, one of the co-founders of PayPal, and Jeff Hammerbacher, a Cloudera founding employee.

“Together is spearheading AI’s ‘Linux moment’ by providing an open ecosystem across compute and best in class foundation models,” Lux Capital’s Brandon Reeves told TechCrunch via email. “Together team is committed to creating a vibrant open ecosystem that allows anyone from individuals to enterprises to participate.”

Together, launched in June 2022, is the brainchild of Vipul Ved Prakash, Ce Zhang, Chris Re and Percy Liang. Prakash previously founded social media search platform Topsy, which was acquired in 2013 by Apple, where he later became a senior director. Zhang is an associate professor of computer science at ETH Zurich, currently on sabbatical and leading research in “decentralized” AI. As for Re, he’s co-founded various startups including SambaNova, which builds hardware and integrated systems for AI. And Liang, a computer science professor at Stanford, directs the university’s Center for Research on Foundation Models (CRFM).

With Together, Prakash, Zhang, Re and Liang are seeking to create open source generative AI models and services that, in their words, “help organizations incorporate AI into their production applications.” To that end, Together is building a cloud platform for running, training and fine-tuning open source models that the co-founders claim will offer scalable compute at “dramatically lower” prices than the dominant vendors (e.g. Google Cloud, AWS, Azure, etc.)

“We believe that generative models are a consequential technology for society and open and decentralized alternatives to closed systems are going to be critical to enable the best outcomes for AI and society,” Prakash told TechCrunch in an email interview. “As enterprises define their generative AI strategies, they’re looking for privacy, transparency, customization and ease of deployment. Current cloud offerings, with closed-source models and data, do not meet their requirements.”

He has a point — insofar as incumbents are feeling the pressure, at least. An internal Google memo leaked earlier in the month implies that the search giant — and its rivals, for that matter — can’t compete against open source AI initiatives over the long run. Meanwhile, OpenAI reportedly is preparing to publicly debut its first open source text-generating AI model amid a proliferation of open source alternatives,

One of Together’s first projects, RedPajama, aims to foster a set of open source generative models including “chat” models along the lines of OpenAI’s ChatGPT. A collaborative work between Together and several groups, including the MILA Québec AI Institute, CRFM and ETH’s data science lab, DS3Lab, RedPajama began with the release of a data set that enables organizations to pre-train models that can be permissively licensed.

Together’s other efforts to date include GPT-JT, a fork of the open source text-generating model GPT-J-6B (released by the research group EleutherAI), and OpenChatKit, an attempt at a ChatGPT equivalent.

“Today, training, fine-tuning or productizing open source generative models is extremely challenging,” Prakash said. “Current solutions require that you have significant expertise in AI and are simultaneously able to manage the large-scale infrastructure needed. The Together platform takes care of both challenges out-of-the-box, with an easy-to-use and accessible solution.”

Just how seamless Together is remains to be seen, though — the platform has yet to launch in GA. And, one might argue, its efforts are a bit duplicative in the context of the broader AI landscape. The number of open source models both from community groups and large labs grows by the day, practically. And while not all are licensed for commercial use, several, like Databricks’ Dolly 2.0, are.

On the AI hardware infrastructure front, besides the big public cloud providers, startups like CoreWeave claim to offer powerful compute for below market rates. There’s even been attempts at building community-powered, free services for running AI text-generating models. (Together intends to follow in the footsteps of these community groups by building a platform, tentatively called the Together Decentralized Cloud, that’ll pool hardware resources including GPUs from volunteers around the internet.)

So what does Together bring to the table? Greater transparency, control and privacy, Prakash argues. It’s a sales pitch not dissimilar to the one made by startup Stability AI, which funnels compute and capital toward open source research while commercializing — and selling services on top of — the various finished products.

“Regulated enterprises will be big customers of open source, as open source models pre-trained on open data sets enable organizations to fully inspect, understand and customize the models to their own applications,” he said. “We believe that the challenges in AI can only be overcome by a global community working together. So we made it our mission to build and steward a self-sustaining, open ecosystem that produces the best AI systems for humanity.”

It’s a lofty goal, to be sure. And it’s early days for Together, which wouldn’t say whether it has any customers at present — much less revenue. But the company is forging ahead, planning to increase the size of its team from 24 employees to around 40 by the end of the year and spend the rest of the seed capital on R&D, infrastructure and product development.

“The Together solution, based on open source generative models, was built on understanding requirements from large organizations and addressing each of these needs, to provide enterprises with the core platform for their generative AI strategy,” Prakash said. “Together is seeing tremendous interest from enterprises looking for greater transparency, control, and privacy.”

Together raises $20M to build open source generative AI models by Kyle Wiggers originally published on TechCrunch

Posted on

The week in AI: Google goes all out at I/O as regulations creep up

Keeping up with an industry as fast-moving as AI is a tall order. So until an AI can do it for you, here’s a handy roundup of the last week’s stories in the world of machine learning, along with notable research and experiments we didn’t cover on their own.

This week, Google dominated the AI news cycle with a range of new products that launched at its annual I/O developer conference. They run the gamut from a code-generating AI meant to compete with GitHub’s Copilot to an AI music generator that turns text prompts into short songs.

A fair number of these tools look to be legitimate labor savers — more than marketing fluff, that’s to say. I’m particularly intrigued by Project Tailwind, a note-taking app that leverages AI to organize, summarize and analyze files from a personal Google Docs folder. But they also expose the limitations and shortcomings of even the best AI technologies today.

Take PaLM 2, for example, Google’s newest large language model (LLM). PaLM 2 will power Google’s updated Bard chat tool, the company’s competitor to OpenAI’s ChatGPT, and function as the foundation model for most of Google’s new AI features. But while PaLM 2 can write code, emails and more, like comparable LLMs, it also responds to questions in toxic and biased ways.

Google’s music generator, too, is fairly limited in what it can accomplish. As I wrote in my hands on, most of the songs I’ve created with MusicLM sound passable at best — and at worst like a four-year-old let loose on a DAW.

There’s been much written about how AI will replace jobs — potentially the equivalent of 300 million full-time jobs, according to a report by Goldman Sachs. In a survey by Harris, 40% of workers familiar with OpenAI’s AI-powered chatbot tool, ChatGPT, are concerned that it’ll replace their jobs entirely.

Google’s AI isn’t the end-all be-all. Indeed, the company’s arguably behind in the AI race. But it’s an undeniable fact that Google employs some of the top AI researchers in the world. And if this is the best they can manage, it’s a testament to the fact that AI is far from a solved problem.

Google I/O 2023 is a wrap — here’s a list of everything announced

Here are the other AI headlines of note from the past few days:

  • Meta brings generative AI to ads: Meta this week announced an AI sandbox, of sorts, for advertisers to help them create alternative copies, background generation through text prompts and image cropping for Facebook or Instagram ads. The company said that the features are available to select advertisers at the moment and will expand access to more advertisers in July.
  • Added context: Anthropic has expanded the context window for Claude — its flagship text-generating AI model, still in preview — from 9,000 tokens to 100,000 tokens. Context window refers to the text the model considers before generating additional text, while tokens represent raw text (e.g., the word “fantastic” would be split into the tokens “fan,” “tas” and “tic”). Historically and even today, poor memory has been an impediment to the usefulness of text-generating AI. But larger context windows could change that.
  • Anthropic touts ‘constitutional AI’: Larger context windows aren’t the Anthropic models’ only differentiator. The company this week detailed “constitutional AI,” its in-house AI training technique that aims to imbue AI systems with “values” defined by a “constitution.” In contrast to other approaches, Anthropic argues that constitutional AI makes the behavior of systems both easier to understand and simpler to adjust as needed.
  • An LLM built for research: The nonprofit Allen Institute for AI Research (AI2) announced that it plans to train a research-focused LLM called Open Language Model, adding to the large and growing open source library. AI2 sees Open Language Model, or OLMo for short, as a platform and not just a model — one that’ll allow the research community to take each component AI2 creates and either use it themselves or seek to improve it.
  • New fund for AI: In other AI2 news, AI2 Incubator, the nonprofit’s AI startup fund, is revving up again at three times its previous size — $30 million versus $10 million. Twenty-one companies have passed through the incubator since 2017, attracting some $160 million in further investment and at least one major acquisition: XNOR, an AI acceleration and efficiency outfit that was subsequently snapped up by Apple for around $200 million.
  • EU intros rules for generative AI: In a series of votes in the European Parliament, MEPs this week backed a raft of amendments to the bloc’s draft AI legislation — including settling on requirements for the so-called foundational models that underpin generative AI technologies like OpenAI’s ChatGPT. The amendments put the onus on providers of foundational models to apply safety checks, data governance measures and risk mitigations prior to putting their models on the market
  • A universal translator: Google is testing a powerful new translation service that redubs video in a new language while also synchronizing the speaker’s lips with words they never spoke. It could be very useful for a lot of reasons, but the company was upfront about the possibility of abuse and the steps taken to prevent it.
  • Automated explanations: It’s often said that LLMs along the lines of OpenAI’s ChatGPT are a black box, and certainly, there’s some truth to that. In an effort to peel back their layers, OpenAI is developing a tool to automatically identify which parts of an LLM are responsible for which of its behaviors. The engineers behind it stress that it’s in the early stages, but the code to run it is available in open source on GitHub as of this week.
  • IBM launches new AI services: At its annual Think conference, IBM announced IBM Watsonx, a new platform that delivers tools to build AI models and provide access to pretrained models for generating computer code, text and more. The company says the launch was motivated by the challenges many businesses still experience in deploying AI within the workplace.

Other machine learnings

Image Credits: Landing AI

Andrew Ng’s new company Landing AI is taking a more intuitive approach to creating computer vision training. Making a model understand what you want to identify in images is pretty painstaking, but their “visual prompting” technique lets you just make a few brush strokes and it figures out your intent from there. Anyone who has to build segmentation models is saying “my god, finally!” Probably a lot of grad students who currently spend hours masking organelles and household objects.

Microsoft has applied diffusion models in a unique and interesting way, essentially using them to generate an action vector instead of an image, having trained it on lots of observed human actions. It’s still very early and diffusion isn’t the obvious solution for this, but as they’re stable and versatile, it’s interesting to see how they can be applied beyond purely visual tasks. Their paper is being presented at ICLR later this year.

Image Credits: Meta

Meta is also pushing the edges of AI with ImageBind, which it claims is the first model that can process and integrate data from six different modalities: images and video, audio, 3D depth data, thermal info, and motion or positional data. This means that in its little machine learning embedding space, an image might be associated with a sound, a 3D shape, and various text descriptions, any one of which could be asked about or used to make a decision. It’s a step towards “general” AI in that it absorbs and associates data more like the brain — but it’s still basic and experimental, so don’t get too excited just yet.

If these proteins touch… what happens?

Everyone got excited about AlphaFold, and for good reason, but really structure is just one small part of the very complex science of proteomics. It’s how those proteins interact that is both important and difficult to predict — but this new PeSTo model from EPFL attempts to do just that. “It focuses on significant atoms and interactions within the protein structure,” said lead developer Lucien Krapp. “It means that this method effectively captures the complex interactions within protein structures to enable an accurate prediction of protein binding interfaces.” Even if it isn’t exact or 100% reliable, not having to start from scratch is super useful for researchers.

The feds are going big on AI. The President even dropped in on a meeting with a bunch of top AI CEOs to say how important getting this right is. Maybe a bunch of corporations aren’t necessarily the right ones to ask, but they’ll at least have some ideas worth considering. But they already have lobbyists, right?

I’m more excited about the new AI research centers popping up with federal funding. Basic research is hugely needed to counterbalance the product-focused work being done by the likes of OpenAI and Google — so when you have AI centers with mandates to investigate things like social science (at CMU), or climate change and agriculture (at U of Minnesota), it feels like green fields (both figuratively and literally). Though I also want to give a little shout out to this Meta research on forestry measurement.

Doing AI together on a big screen — it’s science!

Lots of interesting conversations out there about AI. I thought this interview with UCLA (my alma mater, go Bruins) academics Jacob Foster and Danny Snelson was an interesting one. Here’s a great thought on LLMs to pretend you came up with this weekend when people are talking about AI:

These systems reveal just how formally consistent most writing is. The more generic the formats that these predictive models simulate, the more successful they are. These developments push us to recognize the normative functions of our forms and potentially transform them. After the introduction of photography, which is very good at capturing a representational space, the painterly milieu developed Impressionism, a style that rejected accurate representation altogether to linger with the materiality of paint itself.

Definitely using that!

The week in AI: Google goes all out at I/O as regulations creep up by Kyle Wiggers originally published on TechCrunch

Posted on

Google launches ML Hub to help AI developers train and deploy their models

At its I/O developers conference, Google today announced its new ML Hub, a one-stop destination for developers who want to get more guidance on how to train and deploy their ML models, no matter whether they are in the early stages of their AI career or seasoned professionals.

“We talk about this concept of democratizing machine learning and really making it more accessible, so something that we’re pretty excited about is Google has a bit of a sprawling set of open-source technologies that cover many different assets […] We want to make it much, much easier to understand how they fit together and actually help folks get up and running,” said Alex Spinelli, Google’s VP  of product management for machine learning. The idea here, he said, is to give developers a landing page where they can basically look at what kind of model they want to generate, based on the data they have, and then get step-by-step directions for how to think about deploying those models.

The company is launching this platform with an initial set of toolkits that covers a set of common use cases, with plans to regularly update these and launch new ones in a steady cadence. Some of the early toolkits, for example, can help developers build text classifiers using Keras or take large language models and run them on Android with Keras and TensorFlow Lite.

As Spinelli rightly noted, generative AI may be getting all of the hype right now, but machine learning is a large space that covers a wide range of types of models and technology.

“There’s amazing things going on in computer vision and facial recognition and recommendation systems and relevance ranking of content and those kinds of things — clustering content — all this stuff. We really don’t want to leave anything behind and want to make sure we can actually help developers and researchers have the right set of tools and technologies for their particular use case,” Spinelli noted.

He noted that a lot of the focus here is on open source — and while developers can take these technologies and run them on-premises or in any cloud, these new toolkits will also provide what he called a “glide path into the Google cloud.” But as Spinelli stressed, there is no lock-in here. “There is a fundamental commitment that this is open source that you can use anywhere,” he said.

Read more about Google I/O 2023 on TechCrunch

Google launches ML Hub to help AI developers train and deploy their models by Frederic Lardinois originally published on TechCrunch

Posted on

Google Cloud announces new A3 supercomputer VMs built to power LLMs

As we’ve seen LLMs and generative AI come screaming into our consciousness in recent months, it’s clear that these models take enormous amounts of compute power to train and run. Recognizing this, Google Cloud announced a new A3 supercomputer virtual machine today at Google I/O.

The A3 has been purpose-built to handle the considerable demands of these resource-hungry use cases.

“A3 GPU VMs were purpose-built to deliver the highest-performance training for today’s ML workloads, complete with modern CPU, improved host memory, next-generation NVIDIA GPUs and major network upgrades,” the company wrote in an announcement.

Specifically, the company is arming these machines with NVIDIA’s H100 GPUs and combining that with a specialized data center to derive immense computational power with high throughput and low latency, all at what they suggest is a more reasonable price point than you would typically pay for such a package.

If you’re looking for specs, consider it’s powered by 8 NVIDIA H100 GPUs, 4th Gen Intel Xeon Scalable processors, 2TB of host memory and 3.6 TB/s bisectional bandwidth between the 8 GPUs via NVSwitch and NVLink 4.0, two NVIDIA technologies designed to help maximize throughput between multiple GPUs like the ones in this product.

These machines can provide up to 26 exaFlops of power, which should help improve the time and cost related to training larger machine learning models. What’s more, the workloads on these VMs run in Google’s specialized Jupiter data center networking fabric, which the company describes as, “26,000 highly interconnected GPUs.” This enables “full-bandwidth reconfigurable optical links that can adjust the topology on demand.” The company says this approach should also contribute to bringing down the cost of running these workloads.

The idea is to give customers an enormous amount of power designed to train more demanding workloads, whether that involves complex machine learning models or LLMs running generative AI applications, and to do it in a more cost-effective way.

Google will be offering the A3 in a couple of ways: customers can run it themselves, or if they would prefer, as a managed service where Google handles most of the heavy lifting for them. The do-it-yourself approach involves running the A3 VMs on Google Kubernetes Engine (GKE) and Google Compute Engine (GCE), while the managed service runs the A3 VMs on Vertex AI, the company’s managed machine learning platform.

While the new A3 VMs are being announced today at Google I/O, they are only available for now by signing up for a preview waitlist.

Read more about Google I/O 2023 on TechCrunch

Google Cloud announces new A3 supercomputer VMs built to power LLMs by Ron Miller originally published on TechCrunch

Posted on

Ascend raises $25 million for pre-seed AI startups in the Pacific Northwest

Investing in artificial intelligence (AI) startups is the latest bandwagon VCs are piling onto. But as last year’s crypto experts quickly work to rebrand as AI experts, they’ll have to compete with the VCs who have been investing in the category all along.

Seattle-based Ascend is one of them. Firm founder and solo GP Kirby Winfield has been involved in the AI sector as either a founder or investor since the 90s. Now that seemingly every VC has turned their attention to the category he told TechCrunch he’s glad he’s been in it for so long and therefore will not make some of the mistakes newer entrants will.

“It’s so easy to throw together a vertical AI demo,” Winfield told TechCrunch. “You see a lot of folks who would have been decent SaaS founders, trying to be decent AI founders. I would say it is pretty easy to identify who has actual chops from a technical perspective. We are really fortunate to be investing at this time regardless of the hype.”

Ascend is announcing the close of $25 million for its second fund. Winfield said the firm will invest in pre-seed AI and machine learning (ML) companies largely based in the Pacific Northwest. This continues the firm’s strategy from its first fund which raised $15 million and started deploying in 2019.

Winfield isn’t fully avoiding the hype though. The firm hasn’t always only focused on AI and ML. Ascend’s Fund I also invested in brands and marketplaces too, areas it is stepping away from with this latest batch of capital.

The fund was raised 100% from individuals, Winfield said, and consists of two vehicles: one that raised $22.5 million and another that raised $2.5 million from existing portfolio company founders. Winfield said he was able to raise $21 million in the first month the fund was open before letting it sit open for almost the entirety of 2022 hoping to see some additional funds mosey in, a process he also ran for Fund I.

“I would say that money trickled in a lot more strongly in 2019 when I raised Fund I,” Winfield said. “I couldn’t really think of a good reason to close the fund. We got another $3 million in the door by leaving it open. I don’t overthink these things too much.”

Winfield added that many of the Fund I LPs were happy to reup now that the industry’s notion around investing in AI has changed dramatically since Winfield raised Fund I.

But as every startup is rewriting their marketing to call themselves an AI company, Winfield said he is intentional about the kind of companies he backs. He said he isn’t looking for AI companies necessarily but instead is focused on startups that will utilize the tech to find a better solution.

“AI doesn’t matter,” he said. “What matters is the solution you are selling to your customers. Many founders and investors are getting wrapped around the axle and putting the technology and solution before the benefit.”

Companies from Fund I that fit that bill according to Winfield include Xembly, which uses AI to create a virtual chief of staff, Fabric, which operates as a “headless” e-commerce platform, and WhyLabs, an AI observability platform.

This fund also doubles down on the firm’s focus on companies in the Pacific Northwest, with a particular focus on Seattle. While that might sound limiting for folks who focus on Silicon Valley, Winfield disagrees, citing the talent that comes out of Microsoft and Amazon and the companies that are incubated at the nonprofit Allen Institute for Artificial Intelligence, where Winfield has been the investor in residence for nearly six years.

But no matter his experience and intention, it may still be hard for Winfield to compete with the rapidly growing flock of AI investors. Plus, even if he brings a beneficial background, he doesn’t come with the same deep pockets some of his fellow VCs have — Bessemer just announced they are putting $1 billion of their already raised capital toward the strategy. Plus, we all know how aggressive VCs chasing hype can be.

Xembly founder and CEO Pete Christothoulou said that despite the market’s noise, companies should look to work with VCs like Winfield because while everyone is looking to put money to work in AI, not all support is created equal.

“An AI fund without the right underpinnings is just money,” Christothoulou said. “The money is nice but you want the relationships that the investor can bring. If they can baseline their advice and real technical guidance, that’s where it starts getting really interesting and [Winfield] has a big opportunity.”

Ascend raises $25 million for pre-seed AI startups in the Pacific Northwest by Rebecca Szkutak originally published on TechCrunch