Large Language Models News & Discussions

[bnew]

https://archive.is/Dp5xB

Candle Phi Wasm Demo - a Hugging Face Space by radames

Discover amazing ML apps made by the community

huggingface.co

GitHub - huggingface/candle: Minimalist ML framework for Rust

Minimalist ML framework for Rust. Contribute to huggingface/candle development by creating an account on GitHub.

github.com

 

[bnew]

https://archive.is/wJFaO

Spoiler: thread continued

FunSearch: Making new discoveries in mathematical sciences using Large Language Models

We introduce FunSearch, a method for searching for “functions” written in computer code, and find new solutions in mathematics and computer science. FunSearch works by pairing a pre-trained LLM, whose goal is to provide creative solutions in the form of computer code, with an automated...

deepmind.google

​

FunSearch: Making new discoveries in mathematical sciences using Large Language Models​

Published

14 DECEMBER 2023

Authors

Alhussein Fawzi and Bernardino Romera Paredes

By searching for “functions” written in computer code, FunSearch made the first discoveries in open problems in mathematical sciences using LLMs

Large Language Models (LLMs) are useful assistants - they excel at combining concepts and can read, write and code to help people solve problems. But could they discover entirely new knowledge?

As LLMs have been shown to “hallucinate” factually incorrect information, using them to make verifiably correct discoveries is a challenge. But what if we could harness the creativity of LLMs by identifying and building upon only their very best ideas?

Today, in a paper published in Nature, we introduce FunSearch, a method to search for new solutions in mathematics and computer science. FunSearch works by pairing a pre-trained LLM, whose goal is to provide creative solutions in the form of computer code, with an automated “evaluator”, which guards against hallucinations and incorrect ideas. By iterating back-and-forth between these two components, initial solutions “evolve” into new knowledge. The system searches for “functions” written in computer code; hence the name FunSearch.

This work represents the first time a new discovery has been made for challenging open problems in science or mathematics using LLMs. FunSearch discovered new solutions for the cap set problem, a longstanding open problem in mathematics. In addition, to demonstrate the practical usefulness of FunSearch, we used it to discover more effective algorithms for the “bin-packing” problem, which has ubiquitous applications such as making data centers more efficient.

Scientific progress has always relied on the ability to share new understanding. What makes FunSearch a particularly powerful scientific tool is that it outputs programs that reveal how its solutions are constructed, rather than just what the solutions are. We hope this can inspire further insights in the scientists who use FunSearch, driving a virtuous cycle of improvement and discovery.

Driving discovery through evolution with language models​

FunSearch uses an evolutionary method powered by LLMs, which promotes and develops the highest scoring ideas. These ideas are expressed as computer programs, so that they can be run and evaluated automatically. First, the user writes a description of the problem in the form of code. This description comprises a procedure to evaluate programs, and a seed program used to initialize a pool of programs.

FunSearch is an iterative procedure; at each iteration, the system selects some programs from the current pool of programs, which are fed to an LLM. The LLM creatively builds upon these, and generates new programs, which are automatically evaluated. The best ones are added back to the pool of existing programs, creating a self-improving loop. FunSearch uses Google’s PaLM 2, but it is compatible with other LLMs trained on code.

The FunSearch process. The LLM is shown a selection of the best programs it has generated so far (retrieved from the programs database), and asked to generate an even better one. The programs proposed by the LLM are automatically executed, and evaluated. The best programs are added to the database, for selection in subsequent cycles. The user can at any point retrieve the highest-scoring programs discovered so far.

Discovering new mathematical knowledge and algorithms in different domains is a notoriously difficult task, and largely beyond the power of the most advanced AI systems. To tackle such challenging problems with FunSearch, we introduced multiple key components. Instead of starting from scratch, we start the evolutionary process with common knowledge about the problem, and let FunSearch focus on finding the most critical ideas to achieve new discoveries. In addition, our evolutionary process uses a strategy to improve the diversity of ideas in order to avoid stagnation. Finally, we run the evolutionary process in parallel to improve the system efficiency.

Breaking new ground in mathematics​

We first address the cap set problem, an open challenge, which has vexed mathematicians in multiple research areas for decades. Renowned mathematician Terence Tao once described it as his favorite open question. We collaborated with Jordan Ellenberg, a professor of mathematics at the University of Wisconsin–Madison, and author of an important breakthrough on the cap set problem.

The problem consists of finding the largest set of points (called a cap set) in a high-dimensional grid, where no three points lie on a line. This problem is important because it serves as a model for other problems in extremal combinatorics - the study of how large or small a collection of numbers, graphs or other objects could be. Brute-force computing approaches to this problem don’t work – the number of possibilities to consider quickly becomes greater than the number of atoms in the universe.

FunSearch generated solutions - in the form of programs - that in some settings discovered the largest cap sets ever found. This represents the largest increase in the size of cap sets in the past 20 years. Moreover, FunSearch outperformed state-of-the-art computational solvers, as this problem scales well beyond their current capabilities.

Interactive figure showing the evolution from the seed program (top) to a new higher-scoring function (bottom). Each circle is a program, with its size proportional to the score assigned to it. Only ancestors of the program at the bottom are shown. The corresponding function produced by FunSearch for each node is shown on the right (see full program using this function in the paper).

These results demonstrate that the FunSearch technique can take us beyond established results on hard combinatorial problems, where intuition can be difficult to build. We expect this approach to play a role in new discoveries for similar theoretical problems in combinatorics, and in the future it may open up new possibilities in fields such as communication theory.

FunSearch favors concise and human-interpretable programs​

While discovering new mathematical knowledge is significant in itself, the FunSearch approach offers an additional benefit over traditional computer search techniques. That’s because FunSearch isn’t a black box that merely generates solutions to problems. Instead, it generates programs that describe how those solutions were arrived at. This show-your-working approach is how scientists generally operate, with new discoveries or phenomena explained through the process used to produce them.

FunSearch favors finding solutions represented by highly compact programs - solutions with a low Kolmogorov complexity†. Short programs can describe very large objects, allowing FunSearch to scale to large needle-in-a-haystack problems. Moreover, this makes FunSearch’s program outputs easier for researchers to comprehend. Ellenberg said: “FunSearch offers a completely new mechanism for developing strategies of attack. The solutions generated by FunSearch are far conceptually richer than a mere list of numbers. When I study them, I learn something”.

What’s more, this interpretability of FunSearch’s programs can provide actionable insights to researchers. As we used FunSearch we noticed, for example, intriguing symmetries in the code of some of its high-scoring outputs. This gave us a new insight into the problem, and we used this insight to refine the problem introduced to FunSearch, resulting in even better solutions. We see this as an exemplar for a collaborative procedure between humans and FunSearch across many problems in mathematics.

Left: Inspecting code generated by FunSearch yielded further actionable insights (highlights added by us). Right: The raw “admissible” set constructed using the (much shorter) program on the left.

The solutions generated by FunSearch are far conceptually richer than a mere list of numbers. When I study them, I learn something.

JORDAN ELLENBERG, COLLABORATOR AND PROFESSOR OF MATHEMATICS AT THE UNIVERSITY OF WISCONSIN–MADISON

Addressing a notoriously hard challenge in computing​

Encouraged by our success with the theoretical cap set problem, we decided to explore the flexibility of FunSearch by applying it to an important practical challenge in computer science. The “bin packing” problem looks at how to pack items of different sizes into the smallest number of bins. It sits at the core of many real-world problems, from loading containers with items to allocating compute jobs in data centers to minimize costs.

The online bin-packing problem is typically addressed using algorithmic rules-of-thumb (heuristics) based on human experience. But finding a set of rules for each specific situation - with differing sizes, timing, or capacity – can be challenging. Despite being very different from the cap set problem, setting up FunSearch for this problem was easy. FunSearch delivered an automatically tailored program (adapting to the specifics of the data) that outperformed established heuristics – using fewer bins to pack the same number of items.

Play

Illustrative example of bin packing using existing heuristic – Best-fit heuristic (left), and using a heuristic discovered by FunSearch (right).

Hard combinatorial problems like online bin packing can be tackled using other AI approaches, such as neural networks and reinforcement learning. Such approaches have proven to be effective too, but may also require significant resources to deploy. FunSearch, on the other hand, outputs code that can be easily inspected and deployed, meaning its solutions could potentially be slotted into a variety of real-world industrial systems to bring swift benefits.

LLM-driven discovery for science and beyond​

FunSearch demonstrates that if we safeguard against LLMs’ hallucinations, the power of these models can be harnessed not only to produce new mathematical discoveries, but also to reveal potentially impactful solutions to important real-world problems.

We envision that for many problems in science and industry - longstanding or new - generating effective and tailored algorithms using LLM-driven approaches will become common practice.

Indeed, this is just the beginning. FunSearch will improve as a natural consequence of the wider progress of LLMs, and we will also be working to broaden its capabilities to address a variety of society’s pressing scientific and engineering challenges.

 

[bnew]

Weak-to-strong generalization

We present a new research direction for superalignment, together with promising initial results: can we leverage the generalization properties of deep learning to control strong models with weak supervisors?

openai.com

Weak-to-strong generalization​

Justin Jay Wang ✗ DALL·E

We present a new research direction for superalignment, together with promising initial results: can we leverage the generalization properties of deep learning to control strong models with weak supervisors?.

December 14, 2023

More resources​

• Read paper

Safety & Alignment

A core challenge for aligning future superhuman AI systems (superalignment) is that humans will need to supervise AI systems much smarter than them. We study a simple analogy: can small models supervise large models? We show that we can use a GPT-2-level model to elicit most of GPT-4’s capabilities—close to GPT-3.5-level performance—generalizing correctly even to hard problems where the small model failed. This opens up a new research direction that allows us to directly tackle a central challenge of aligning future superhuman models while making iterative empirical progress today.

The superalignment problem​

We believe superintelligence—AI vastly smarter than humans—could be developed within the next ten years. However, we still do not know how to reliably steer and control superhuman AI systems. Solving this problem is essential for ensuring that even the most advanced AI systems in the future remain safe and beneficial to humanity.

We formed the Superalignment team earlier this year to solve this problem of superintelligence alignment. Today, we are releasing the team’s first paper, which introduces a new research direction for empirically aligning superhuman models.

Current alignment methods, such as reinforcement learning from human feedback (RLHF), rely on human supervision. However, future AI systems will be capable of extremely complex and creative behaviors that will make it hard for humans to reliably supervise them. For example, superhuman models may be able to write millions of lines of novel—and potentially dangerous—computer code that would be very hard even for expert humans to understand.

Relative to superhuman AI models, humans will be “weak supervisors.” This is a core challenge for AGI alignment: how can weak supervisors trust and control substantially stronger models?

Our setup​

To make progress on this core challenge, we propose an analogy we can empirically study today: can we use a smaller (less capable) model to supervise a larger (more capable) model?

A simple analogy for superalignment: In traditional machine learning (ML), humans supervise AI systems weaker than themselves (left). To align superintelligence, humans will instead need to supervise AI systems smarter than them (center). We cannot directly study this problem today, but we can study a simple analogy: can small models supervise larger models (right)?

Naively, we might not expect a strong model to perform better than the weak supervisor that provides its training signal—it may simply learn to imitate all the errors the weak supervisor makes. On the other hand, strong pretrained models have excellent raw capabilities—we don't need to teach them new tasks from scratch, we just need to elicit their latent knowledge. The critical question is then: will the strong model generalize according to the weak supervisor's underlying intent—leveraging its full capabilities to solve the task even on difficult problems where the weak supervisor can only provide incomplete or flawed training labels?

Our results​

Typical weak-to-strong generalization across NLP benchmarks: We use a GPT-2-level model as a weak supervisor to finetune GPT-4.

We can significantly improve generalization in many settings. We use a simple method that encourages the strong model to be more confident—including confidently disagreeing with the weak supervisor if necessary. When we supervise GPT-4 with a GPT-2-level model using this method on NLP tasks, the resulting model typically performs somewhere between GPT-3 and GPT-3.5. We are able to recover much of GPT-4’s capabilities with only much weaker supervision.

This method is a proof of concept with important limitations; for example, it still doesn’t work on ChatGPT preference data. However, we also find signs of life with other approaches, such as optimal early stopping and bootstrapping from small to intermediate to large models.

Collectively, our results suggest that (1) naive human supervision—such as reinforcement learning from human feedback (RLHF)—could scale poorly to superhuman models without further work, but (2) it is feasible to substantially improve weak-to-strong generalization.

Research opportunities​

There are still important disanalogies between our current empirical setup and the ultimate problem of aligning superhuman models. For example, it may be easier for future models to imitate weak human errors than for current strong models to imitate current weak model errors, which could make generalization harder in the future.

Nevertheless, we believe our setup captures some key difficulties of aligning future superhuman models, enabling us to start making empirical progress on this problem today. There are many promising directions for future work, including fixing the disanalogies in our setup, developing better scalable methods, and advancing our scientific understanding of when and how we should expect good weak-to-strong generalization.

We believe this is an exciting opportunity for the ML research community to make progress on alignment. To kickstart more research in this area,

• We are releasing open source code to make it easy to get started with weak-to-strong generalization experiments today.
• We are launching a $10 million grants program for graduate students, academics, and other researchers to work on superhuman AI alignment broadly. We’re especially excited to support research related to weak-to-strong generalization.

Figuring out how to align future superhuman AI systems to be safe has never been more important, and it is now easier than ever to make empirical progress on this problem. We are excited to see what breakthroughs researchers discover.

 

[bnew]

Weak-to-strong generalization

We present a new research direction for superalignment, together with promising initial results: can we leverage the generalization properties of deep learning to control strong models with weak supervisors?

openai.com

Weak-to-strong generalization​

Justin Jay Wang ✗ DALL·E

We present a new research direction for superalignment, together with promising initial results: can we leverage the generalization properties of deep learning to control strong models with weak supervisors?.

December 14, 2023

More resources​

• Read paper

Safety & Alignment

A core challenge for aligning future superhuman AI systems (superalignment) is that humans will need to supervise AI systems much smarter than them. We study a simple analogy: can small models supervise large models? We show that we can use a GPT-2-level model to elicit most of GPT-4’s capabilities—close to GPT-3.5-level performance—generalizing correctly even to hard problems where the small model failed. This opens up a new research direction that allows us to directly tackle a central challenge of aligning future superhuman models while making iterative empirical progress today.

The superalignment problem​

We believe superintelligence—AI vastly smarter than humans—could be developed within the next ten years. However, we still do not know how to reliably steer and control superhuman AI systems. Solving this problem is essential for ensuring that even the most advanced AI systems in the future remain safe and beneficial to humanity.

We formed the Superalignment team earlier this year to solve this problem of superintelligence alignment. Today, we are releasing the team’s first paper, which introduces a new research direction for empirically aligning superhuman models.

Current alignment methods, such as reinforcement learning from human feedback (RLHF), rely on human supervision. However, future AI systems will be capable of extremely complex and creative behaviors that will make it hard for humans to reliably supervise them. For example, superhuman models may be able to write millions of lines of novel—and potentially dangerous—computer code that would be very hard even for expert humans to understand.

Relative to superhuman AI models, humans will be “weak supervisors.” This is a core challenge for AGI alignment: how can weak supervisors trust and control substantially stronger models?

Our setup​

To make progress on this core challenge, we propose an analogy we can empirically study today: can we use a smaller (less capable) model to supervise a larger (more capable) model?

A simple analogy for superalignment: In traditional machine learning (ML), humans supervise AI systems weaker than themselves (left). To align superintelligence, humans will instead need to supervise AI systems smarter than them (center). We cannot directly study this problem today, but we can study a simple analogy: can small models supervise larger models (right)?

Naively, we might not expect a strong model to perform better than the weak supervisor that provides its training signal—it may simply learn to imitate all the errors the weak supervisor makes. On the other hand, strong pretrained models have excellent raw capabilities—we don't need to teach them new tasks from scratch, we just need to elicit their latent knowledge. The critical question is then: will the strong model generalize according to the weak supervisor's underlying intent—leveraging its full capabilities to solve the task even on difficult problems where the weak supervisor can only provide incomplete or flawed training labels?

Our results​

Typical weak-to-strong generalization across NLP benchmarks: We use a GPT-2-level model as a weak supervisor to finetune GPT-4.

We can significantly improve generalization in many settings. We use a simple method that encourages the strong model to be more confident—including confidently disagreeing with the weak supervisor if necessary. When we supervise GPT-4 with a GPT-2-level model using this method on NLP tasks, the resulting model typically performs somewhere between GPT-3 and GPT-3.5. We are able to recover much of GPT-4’s capabilities with only much weaker supervision.

This method is a proof of concept with important limitations; for example, it still doesn’t work on ChatGPT preference data. However, we also find signs of life with other approaches, such as optimal early stopping and bootstrapping from small to intermediate to large models.

Collectively, our results suggest that (1) naive human supervision—such as reinforcement learning from human feedback (RLHF)—could scale poorly to superhuman models without further work, but (2) it is feasible to substantially improve weak-to-strong generalization.

Research opportunities​

There are still important disanalogies between our current empirical setup and the ultimate problem of aligning superhuman models. For example, it may be easier for future models to imitate weak human errors than for current strong models to imitate current weak model errors, which could make generalization harder in the future.

Nevertheless, we believe our setup captures some key difficulties of aligning future superhuman models, enabling us to start making empirical progress on this problem today. There are many promising directions for future work, including fixing the disanalogies in our setup, developing better scalable methods, and advancing our scientific understanding of when and how we should expect good weak-to-strong generalization.

We believe this is an exciting opportunity for the ML research community to make progress on alignment. To kickstart more research in this area,

• We are releasing open source code to make it easy to get started with weak-to-strong generalization experiments today.
• We are launching a $10 million grants program for graduate students, academics, and other researchers to work on superhuman AI alignment broadly. We’re especially excited to support research related to weak-to-strong generalization.

Figuring out how to align future superhuman AI systems to be safe has never been more important, and it is now easier than ever to make empirical progress on this problem. We are excited to see what breakthroughs researchers discover.

 

[bnew]

News Execs: Google’s AI Integration Could ‘Kill Journalism’

News publishers are reportedly alarmed by what Google’s artificial intelligence (AI) tool means for their business.

www.pymnts.com

News Execs: Google’s AI Integration Could ‘Kill Journalism’​

BY PYMNTS | DECEMBER 14, 2023

News publishers are reportedly alarmed by what Google’s AI means for their business.

As the Wall Street Journal (WSJ) reported Thursday (Dec. 14), Google’s move to integrate its “Search Generative Experience” artificial intelligence (AI) tool has underlined the dangers for media companies of depending on the tech giant to bring their stories to readers.

The report notes that news outlets are already suffering from a drop in traffic from social media sites, with both Facebook and Twitter moving away from news distribution. Google’s AI-powered search could make things worse, as Google generates almost 40% of publishers’ traffic, the report said, citing a WSJ analysis of data from Similarweb.

The WSJ also cites findings from a task force at the news magazine The Atlantic, which gets about 40% of its web traffic from Google searches.

That task force studied what would happen if Google integrated AI into search, and found that 75% of the time, the AI-powered search would deliver a complete answer to a user’s question and deprive the Atlantic of that potential traffic.

“AI and large language models have the potential to destroy journalism and media brands as we know them,” said Mathias Dopfner, chairman and CEO of Axel Springer.

His company, which owns Business Insider and Politico, announced Wednesday (Dec. 13) a deal to license content to OpenAI.

“We want to explore the opportunities of AI empowered journalism — to bring quality, societal relevance and the business model of journalism to the next level,” Dopfer said in announcing that partnership.

According to the WSJ report, Google says its AI product is still being developed, though publishers say they have seen enough to determine they’ll lose 20% to 40% of Google-generated traffic. Google has said it is prioritizing sending traffic to publishers.

However, news publishers aren’t just concerned about AI’s impact on traffic. Earlier this year, a group of major news media organizations called for the regulation of AI model operators on their use of copyrighted materials.

Industry bodies such as the News Media Alliance and the European Publishers’ Council are pushing for a framework that lets them “collectively negotiate” with AI model operators regarding the use of their intellectual property, according to a letter published in August.

“Generative AI and large language models … disseminate that content and information to their users, often without any consideration of, remuneration to, or attribution to the original creators,” the letter said. “Such practices undermine the media industry’s core business models.”

 

[bnew]

 

[Matt504]

 

[greenvale]

So I was messing around with Google's AI studio last night. I don't understand what they mean by building an app though. Everything seems to be some variant of a different chat bot. Can any of you brehs explain the types of stuff I can build? Sorry if this is too obtuse of a question

 

[bnew]

https://archive.is/qU59o

 

[bnew]

Intel unveils new AI chip to compete with Nvidia and AMD

Intel announced new details about PC, server, and AI chips on Thursday, including Gaudi3, an AI chip for generative AI software that will compete with Nvidia.

www.cnbc.com

TECH

Intel unveils new AI chip to compete with Nvidia and AMD​

PUBLISHED THU, DEC 14 2023
11:20 AM EST
UPDATED 44 MIN AGO

Kif Leswing@KIFLESWING

KEY POINTS

• Intel unveiled new computer chips on Thursday, including Gaudi3, a chip for generative AI software.
• Intel also announced Core Ultra chips, designed for Windows laptops and PCs, and new fifth-generation Xeon server chips.
• Intel’s server and PC processors include specialized AI parts called NPUs that can be used to run AI programs faster.
  

In this article

• NVDA+8.60 (+1.78%)
• AMD+0.48 (+0.35%)
• INTC+1.34 (+2.98%)
  

Patrick Gelsinger, chief executive officer of Intel Corp., speaks during the Intel AI Everywhere launch event in New York, US, on Thursday, Dec. 14, 2023.

Victor J. Blue | Bloomberg | Getty Images

Intel unveiled new computer chips on Thursday, including Gaudi3, an artificial intelligence chip for generative AI software. Gaudi3 will launch next year and will compete with rival chips from Nvidia and AMD that power big and power-hungry AI models.

The most prominent AI models, like OpenAI’s ChatGPT, run on Nvidia GPUs in the cloud. It’s one reason Nvidia stock has been up nearly 230% year to date while Intel shares have risen 68%. And it’s why companies like AMD and, now Intel, have announced chips that they hope will attract AI companies away from Nvidia’s dominant position in the market.

Shares of Intel were up 1% on Thursday.

While the company was light on details, Gaudi3 will compete with Nvidia’s H100, the main choice among companies that build huge farms of the chips to power AI applications, and AMD’s forthcoming MI300X, when it starts shipping to customers in 2024.

Intel has been building Gaudi chips since 2019, when it bought a chip developer called Habana Labs.

An Intel Gaudi2 AI accelerator during the Intel AI Everywhere launch event in New York, US, on Thursday, Dec. 14, 2023. Intel Corp. announced new chips for PCs and data centers that the company hopes will give it a bigger slice of the booming market for artificial intelligence hardware. Photographer: Victor J. Blue/Bloomberg via Getty Images

Bloomberg | Bloomberg | Getty Images

“We’ve been seeing the excitement with generative AI, the star of the show for 2023,” Intel CEO Pat Gelsinger said at a launch event in New York where he announced Gaudi3 along other chips focused on AI applications.

“We think the AI PC will be the star of the show for the upcoming year,” Gelsinger added. And that’s where Intel’s new Core Ultra processors, also announced on Thursday, will come into play.

Intel Core Ultra and new Xeon chips​

Intel also announced Core Ultra chips, designed for Windows laptops and PCs, and new fifth-generation Xeon server chips. Both include a specialized AI part called an NPU that can be used to run AI programs faster.

It’s the latest sign that traditional processor makers, including Intel rivals AMD and Qualcomm, are reorienting their products lines around and alerting investors to the possibility of AI models leading to surging demand for their chips.

The Core Ultra won’t provide the same kind of power to run a chatbot like ChatGPT without an internet connection, but can handle smaller tasks. For example, Intel said, Zoom runs its background-blurring feature on its chips. They’re built using the company’s 7-nanometer process, which is more power efficient than earlier chips.

But, importantly, the 7-nanometer chips show Gelsinger’s strategy to catch up to Taiwan Semiconductor Manufacturing Co. in terms of chip manufacturing prowess by 2026 hasn’t fallen behind.

Core Ultra chips also include more powerful gaming capabilities and the added graphics muscle can help programs like Adobe Premier run more than 40% faster. The lineup launched in laptops that hit stores on Thursday.

Finally, Intel’s fifth-generation Xeon processors power servers deployed by large organizations like cloud companies. Intel didn’t share pricing, but the previous Xeon cost thousands of dollars. Intel’s Xeon processors are often paired with Nvidia GPUs in the systems that are used for training and deploying generative AI. In some systems, eight GPUs are paired to one or two Xeon CPUs.

Intel said the latest Xeon processor will be particularly good for inferencing, or the process of deploying an AI model, which is less power hungry than the training process.

 

[bnew]

Ola Founder’s AI Startup Launches Indian Large Language Model

Krutrim, an artificial intelligence startup founded by serial entrepreneur Bhavish Aggarwal, has launched India’s first multilingual large language model, which can generate text in 10 Indian languages.

www.bloomberg.com

Ola Founder’s AI Startup Launches Indian Large Language Model​

• Krutrim’s AI model can generate text in 10 Indian languages
• Company will also develop data centers, super computers for AI
  

Bhavish Aggarwal
Photographer: Aparna Jayakumar/Bloomberg

By Saritha Rai
December 15, 2023 at 8:13 AM EST

Krutrim, an artificial intelligence startup founded by serial entrepreneur Bhavish Aggarwal, has launched India’s first multilingual large language model, which can generate text in 10 Indian languages.

“Our LLM is voice-enabled, and able to understand several languages and even as a mix of languages such as Hinglish — Hindi and English,” Aggarwal said at the Friday event in Bangalore, hosted on the campus of Ola Group, where he’s chief executive officer. “It’s uniquely Indian.”

Krutrim, which translates to “artificial” in Sanskrit, is also developing data centers and will ultimately aim to create servers and super computers for the AI ecosystem. The chatbot will be available in a beta version in January. Prototypes of the servers will be out in mid 2024 and production will start by the end of 2025, the startup said in a statement.

A clutch of Indian startups and academic groups are racing to build large language models in Indian languages, so called Indic LLMs, after the launch of OpenAI’s ChatGPT a year ago. Countries are hoping to build their own competing AI systems, rather than relying on technology from the US or China. In Europe, investors are pouring cash into France’s Mistral AI, now valued at $2 billion after being founded earlier this year. The United Arab Emirates touts its Falcon model, which is backed by an Abu Dhabi government research institute.

India, with 1.4 billion people, is focusing on building smaller, more cost efficient AI systems. Generative AI startup Sarvam, which built its system using available open-source models, launched OpenHathi, its first open-source Hindi LLM earlier this week. The announcement came days after it had raised $41 million in an investment from Lightspeed Venture Partners, billionaire Vinod Khosla and others.

At the event, Aggarwal prompted the open-source Krutrim model to welcome guests in English, write a poem in Tamil, compose an ode to monsoons in Bengali and produce software code. “The AI models known around the world are trained largely in English,” he said. “They cannot capture our culture, language and ethos.”

The company is also focused on developing chips, including a “multiple chiplet” strategy that it said will cut costs and make data center design, he said.

Krutrim — which is widely deployed within the Ola Group ride-hailing company to aid voice chat, sales calls and customer support emails — also plans to roll out an enterprise model called Krutrim Pro next quarter. Aggarwal said he uses the software to write performance reviews for his team and compose job descriptions for hiring.

 

[bnew]

Intel CEO attacks Nvidia on AI: 'The entire industry is motivated to eliminate the CUDA market'

Reframing the AI framework, away from CUDA and toward more open standards.

www.tomshardware.com

Intel CEO attacks Nvidia on AI: 'The entire industry is motivated to eliminate the CUDA market'​

News
By Andrew E. Freedman
published 1 day ago

Reframing the AI framework, away from CUDA and toward more open standards.

(Image credit: Tom's Hardware)

Intel CEO Pat Gelsinger came out swinging at Nvidia's CUDA technology, claiming that inference technology will be more important than training for AI as he launched Intel Core Ultra and 5th Gen Xeon datacenter chips in an event here in New York City. Taking questions at the NASDAQ, Gelsinger suggested that Nvidia’s CUDA dominance in training wouldn't last forever.

"You know, the entire industry is motivated to eliminate the CUDA market," Gelsinger said. He cited examples such as MLIR, Google, and OpenAI, suggesting that they are moving to a "Pythonic programming layer" to make AI training more open.

"We think of the CUDA moat as shallow and small," Gelsinger went on. "Because the industry is motivated to bring a broader set of technologies for broad training, innovation, data science, et cetera."

But Intel isn't relying just on training. Instead, it thinks inference is the way to go.

"As inferencing occurs, hey, once you've trained the model… There is no CUDA dependency," Gelsinger continued. "It's all about, can you run that model well?" He suggested that with Gaudi 3, shown on stage for the first time, that Intel will be up to the challenge, and will be able to do it as well with Xeon and edge PCs. Not that Intel won't compete in training, but "fundamentally, the inference market is where the game will be at," Gelsinger said.

(Image credit: Tom's Hardware)

He also took the opportunity to push OpenVINO, the standard that Intel has gathered around for its AI efforts, and predicted a world of mixed computing, some that occurs in the cloud, and others that happen on your PC.

Sandra Rivera, executive vice president and general manager of the Data Center and AI Group at Intel, added that Intel's scale from the data center to the PC may make it a partner of choice, as it can produce at volume.

"We're going to compete three ways for 100% of the datacenter AI TAM." Gelsinger said, tacking onto Rivera's comment. "With our leadership CEOs, leadership accelerators, and as a foundry. Every one of those internal opportunities is available to us: The TPUs, the inferentias, the trainiums, et cetera. We're going to pursue all of those. And we're going to pursue every commercial opportunity as well, with NVIDIA, with AMD, et cetera. We're going to be a foundry player."

It's a bold strategy, and Gelsinger appeared confident as he led his team through presentations today. Can he truly take on CUDA? Only time will tell as applications for the chips Intel launched today — and that his competitors are also working on — become more widespread.

 

[bnew]

Ola's Bhavish Aggarwal Unveils 'Krutrim’ LLM for AI Development in Indian Languages

Ola's founder Bhavish Aggarwal introduced Krutrim, a LLM that can write stories, poems, and more in multiple Indian languages.

mpost.io

Ola’s Bhavish Aggarwal Unveils ‘Krutrim’ LLM for AI Development in Indian Languages​

Share this article
by Kumar Gandharv

Published: December 15, 2023 at 7:25 am Updated: December 15, 2023 at 7:25 am

by Victor Dey

Edited and fact-checked: December 15, 2023 at 7:25 am

IN BRIEF

Ola’s founder Bhavish Aggarwal introduced Krutrim – an LLM that can write stories, poems, and more in multiple Indian languages.

Indian artificial intelligence company Krutrim Si Designs introduced Krutrim — a large language model (LLM) that can write stories, poems and more in multiple Indian languages.



The company, Krutrim Si Designs was launched in April 2023 and is spearheaded by Bhavish Aggarwal, founder and chief executive officer of Indian ride-hailing company Ola.

Aggarwal introduced Krutrim – meaning artificial in the Sanskrit language – during a live-streamed hybrid event on YouTube in two sizes.

A base model named Krutrim was trained on two trillion pieces of talk (tokens) and datasets. It can understand 20 Indian languages and can even respond in 10 of those languages, including Hindi, Marathi, Telugu, Kannada and more.

Additionally, Krutrim Pro — another multimodal AI model, is slated for release in 2024 and promises more sophisticated problem-solving capabilities and task execution.

According to the announcement, the Krutrim AI model is designed to cater to the unique needs of the Indian context and represents a leap towards technological self-sufficiency and cultural assimilation.

“Today, all AI models, called LLM, are trained largely in English, but language is not just text. Language is also the vehicle for cultural values, context and ethos, and its current AI models just can’t capture India’s culture, knowledge, and aspirations given our multicultural, multilingual heritage,” said Aggarwal.

He further added that the model’s mission is to shape a culturally expressive India, steering towards an AI-first economy, and challenging global paradigms.

Ravi Jain, Ola’s Chief Marketing Officer, shed light on the model’s architecture, stating that Krutrim boasts numerous innovations and represents the largest data set used in training. Positioned as a model that strikes the right balance between performance and price, it is expected to power a myriad of database applications.

We're ready! India's own AI, the journey begins today! Join us here. https://t.co/eyBeiOTKgB



— Bhavish Aggarwal (@bhash) December 15, 2023​

Indian Players Race for LLMs in Local Language​

The Krutrim team claims that the AI model surpasses GPT-4 in Indic language support. The model’s ability to understand and generate output in 20 Indian languages, outperforming its global counterparts, positions it as a frontrunner in the AI race.

To make Krutrim accessible to the public, Ola announced the launch of Krutrim AI model website. Interested users can register their interest using their phone numbers, with the promise of being informed as soon as the chatbot becomes available.

Aggarwal had earlier emphasized that Krutrim would be an AI enterprise tailored for Indian customers, leveraging the rich data resources available within the country.

Recently, another Indian Generative AI startup Sarvam AI released OpenHathi-Hi-v0.1, the first Hindi large language model (LLM) in the OpenHathi series, a week after raising $41 million in its Series A funding round led by Lightspeed Ventures.

According to Sarvam AI, the goal of the OpenHathi series is to contribute to the development of open models and datasets to promote innovation in Indian language AI and is collaborating with academic partners at AI4Bharat, who have provided language resources and benchmarks for this initiative.

India also unveiled BharatGPT, an LLM-based solution created in collaboration with the Indian conversational AI startup, CoRover.ai. Positioned as a response to current generative AI and LLMs, BharatGPT aims to address a significant void in the market by offering an indigenous solution. It is available in more than 12 Indian languages and supports interactions across video, voice, and text platforms.

It will be interesting to see how such local Indian LLMs will disrupt industries, as it may open doors for higher interest in this space worldwide.

 

[bnew]

https://www.washingtonpost.com/technology/2023/12/14/grok-grimes-curio-toy/

Grimes is working on an interactive AI toy for kids. Meet Grok.​

Grok is the first in a line of toys from start-up Curio intended to allow kids to converse with an AI chatbot​

By Taylor Lorenz

Updated December 14, 2023 at 4:26 p.m. EST|Published December 14, 2023 at 7:00 a.m. EST

Grok is an AI-powered plush toy in the shape of a rocket that can fully interact and converse with your child. (Curio)


For decades, movies, TV shows and books have imagined a world where human beings form deep personal bonds with creations powered by artificial intelligence. Movies such as “Her” and “The Creator” present a future where there is almost no distinction between human-to-human friendships and human-to-AI friendships.

A glimpse toward this future is beginning to emerge in products like Grok, an AI-powered plush toy in the shape of a rocket that can converse with your child. Grok is the first product from a Silicon Valley start-up called Curio that is leveraging Open AI’s technology on a line of toys Curio’s founders say will be capable of long-running, fully interactive conversation, allowing a child to view it almost as a peer or friend.

Canadian musician Claire Boucher, known as Grimes and the mother of three of Elon Musk’s children, is an investor in and adviser to the company, and she will provide the toy’s voice.

“Every [change] in technology unlocks new forms of entertainment,” said Sam Eaton, president and chief toy maker at Curio, who was previously an engineer at Roblox, the gaming platform. “It’s a different level of immersion for playtime.”

“It’s a new hardware medium,” said Misha Sallee, Curio’s chief executive.

Academics and investors said that it’s too soon to assess the likely impact of such toys, but the AI toy market is expected to explode in coming years. A recent report by Contrive Datum Insights, a market research firm, predicted that the AI toy market will be worth $35.11 billion by 2030, and several AI-enabled robots for children were on display at New York’s Jacob K. Javits Convention Center during its annual Toy Fair in September.

“Parents always want to buy their kids things that will help them learn and grow, which will make smart AI toys even more popular,” Contrive Datum Insights said in a LinkedIn post about its research.

Given OpenAI’s penchant for providing incorrect or disconcerting information, an AI-enabled toy’s utility as an educational device remains unproven. For now, however, Curio’s founders aren’t pitching it as an educational toy, but rather as an antidote to children’s reliance on screens for entertainment. “I really feel like this is also the first step towards also sort of reducing screen time as much as humanly possible,” Grimes said in a written interview with the founders and the AI theorist Roon.

Eaton said he thinks a lot about the screen-addled world his own children are growing up in, where so much of kids’ entertainment is centered around passive consumption. That entertainment is only getting more addictive, he said. “Finally,” Eaton said, “technology is allowing us to move back into our own reality instead of descending into a digital reality.”

“Our big vision is increasing imagination levels,” Sallee said. “Replacing more-toxic forms of technology is a North Star of sorts for us.”

As the product evolves, Curio’s founders want to give parents even more control over Grok’s conversations. “If this is going to be a guest in the house, we want parents to have influence over what the toy believes,” Eaton said. “Maybe your family is vegetarian or religious and you want to influence [conversation around those topics]. Or maybe your kid has certain interests, like dinosaurs, and you want to put them in there.”

From left, Curio CEO Misha Sallee Grimes, and Curio’s chief toy maker, Sam Eaton. (Curio)


Unlike previous talking toys such as Teddy Ruxpin, Talk to Me Barbie or Furbies, which simply play prerecorded lines, Eaton said his plans for Curio’s toys is for them to “have a degree of some kind of pseudo consciousness.”

Eaton said Grok will come with an app for parents that will provide full transcripts of any conversation the child has with the toy. Parents also will be able to make certain words or topics off limits. The toy won’t collect or store any voice data, and storage of the transcripts will comply with the federal Children’s Online Privacy Protection Act, the company says. Parents can request any transcripts be deleted at any time.

“I really think it’s important to let parents know that they do have full control over it and we’re not hiding anything,” Eaton said.

Sallee said that the toy was designed with Grimes’s children in mind and that they have a friendly relationship with it. “The toy was designed for X and the other kids,” she said, referring to the son of Grimes and Musk, X Æ A-Xii, “but X primarily because he’s of age where he can actually talk to the toy and it can talk back more effectively.”

But the toy has no relationship with Musk’s AI start-up, which also is called Grok. Curio holds the trademark on the name, and the two AI products are totally unaffiliated, Curio says. The name Grok was devised by Grimes and the Curio team, who said the word was a shortening of the word Grocket, which was coined because Grimes’ children are exposed to a lot of rockets through their father’s ownership of SpaceX.

Grok is available for preorder now for $99. Those who order by Dec. 17 will receive a golden ticket in the mail before Christmas. The product itself will ship early next year.

The current version requires a WiFi connection, though Eaton’s hope is that one day the technology will advance so that the toy itself can contain all the hardware and software needed to be interactive; incorporating such technology now would push its price to several thousands of dollars, making it inaccessible to most parents.

Curio also envisions Grok as an assistive technology for parenting. For instance, parents may be able to create a prompt for bedtime and Grok might go into bedtime mode, in which the toy redirects all conversations with the child toward bedtime, or Grok gets sleepy himself.

“We’ve heard some people’s concerns like, ‘Could this replace parents?’” Eaton said. “But it’s the opposite. Kids are already zoned out on YouTube or tablets. This toy, you’re talking to it and it’s talking back. You’re imagining and it’s challenging you.”

The Curio founders said they imagine a future where AI toys will be able to interact and play off each other, like a real life Toy Story.

Curio was incorporated earlier this year and now has four full-time employees and several contractors across marketing and manufacturing, in addition to Curio’s two founders.

The company has attracted seed-stage investment from Grimes as well as technologists including Daniel Gross and Nat Friedman, the former CEO of GitHub. Eaton and Sallee connected with Grimes through a mutual friend, after she expressed interest in the idea of a sentient toy on X.

They also hope to open the technology to more collaborators and partners. “The voice box is the technology, it’s separate from the skin,” Eaton said. “So people could create any kind of plush skin for it. The thinking is we’d open it up and let others build on it to make their own characters.”

“The way Nintendo brought a bit more creativity and art to video games and Pixar did to animation, that’s what we want to do with toys,” he added.

 

[bnew]

Agility is using large language models to communicate with its humanoid robots​

Brian Heater@bheater / 2:46 PM EST•December 14, 2023

Image Credits: Amazon

I’ve spent much of the past year discussing generative AI and large language models with robotics experts. It’s become increasingly clear that these sorts of technologies are primed to revolutionize the way robots communicate, learn, look and are programmed.

Accordingly, a number of top universities, research labs and companies are exploring the best methods for leveraging these artificial intelligence platforms. Well-funded Oregon-based startup Agility has been playing around with the tech for a while now using its bipedal robot, Digit.

Today, the company is showcasing some of that work in a short video shared through its social channels.



“[W]e were curious to see what can be achieved by integrating this technology into Digit,” the company notes. “A physical embodiment of artificial intelligence created a demo space with a series of numbered towers of several heights, as well as three boxes with multiple defining characteristics. Digit was given information about this environment, but was not given any specific information about its tasks, just natural language commands of varying complexity to see if it can execute them.”

In the video example, Digit is told to pick up a box the color of “Darth Vader’s lightsaber” and move it to the tallest tower. The process isn’t instantaneous, but rather slow and deliberate, as one might expect from an early-stage demo. The robot does, however, execute the task as described.

Agility notes, “Our innovation team developed this interactive demo to show how LLMs could make our robots more versatile and faster to deploy. The demo enables people to talk to Digit in natural language and ask it to do tasks, giving a glimpse at the future.”

Natural language communication has been a key potential application for this technology, along with the ability to program systems via low- and no-code technologies.

During my Disrupt panel, Gill Pratt described how the Toyota Research Institute is using generative AI to accelerate robotic learning:

We have figured out how to do something, which is use modern generative AI techniques that enable human demonstration of both position and force to essentially teach a robot from just a handful of examples. The code is not changed at all. What this is based on is something called diffusion policy. It’s work that we did in collaboration with Columbia and MIT. We’ve taught 60 different skills so far.

MIT CSAIL’s Daniela Rus also recently told me, “It turns out that generative AI can be quite powerful for solving even motion planning problems. You can get much faster solutions and much more fluid and human-like solutions for control than with model predictive solutions. I think that’s very powerful, because the robots of the future will be much less roboticized. They will be much more fluid and human-like in their motions.”

The potential applications here are broad and exciting — and Digit, as an advanced commercially available robotic system that is being piloted at Amazon fulfillment centers and other real-world locations, seems like a prime candidate. If robotics are going to work alongside humans, they’ll need to learn to listen to them, as well.