• Posts
  • Meta's Shocking New Research | Self-Rewarding Language Models

Meta's Shocking New Research | Self-Rewarding Language Models

PLUS: Google DeepMind Scientists Startup, Elon Musk's xAI Startup Funding and more.


Meta's Shocking New Research | Self-Rewarding Language Models

AI can train itself using its own created data, kind of like how we humans learn from thinking things through or dreaming. The big question is: Can this self-made data help AI get smarter?

AI is now learning not just from us humans but also by teaming up with other AIs. This process is like AI teaching AI, leading to what they call synthetic intelligences. This approach may lead to AI systems surpassing human limitations, contributing to the continuous advancement of AI capabilities.

Google DeepMind Scientists in Talks to Leave and Form AI Startup

Two Google DeepMind scientists are eyeing a big move. They're chatting with investors about starting their own AI gig in Paris, aiming for over $220 million in funding. Laurent Sifre and Karl Tuyls, both big names in AI, are behind this. They're thinking of building a fresh AI model. 

These guys are no rookies; Sifre helped make AI history with a computer beating Go masters, and Tuyls is a whiz in AI game theory. Their plan's drawing major investor attention, especially in France where tech startups are hot. This move is separate from another London-based biz called Holistic AI.

Elon Musk’s AI startup that hopes to leapfrog OpenAI is halfway to its goal of landing $1 billion in funding

Elon Musk's AI startup, xAI, is halfway to its $1 billion funding goal, bagging $500 million so far. The company, aiming to outdo OpenAI (which Musk co-founded but left), is now valued between $15 and $20 billion, though that could change. xAI's star product, a chatbot named Grok, uses social media data from X (previously Twitter, also owned by Musk) for up-to-date info. 

Big names from Musk's Twitter buyout, like Larry Ellison and Saudi Prince Alwaleed bin Talal, are likely to back xAI too. Investors in X get a chance to put cash into xAI, proportional to their X investment. In the next few weeks, Musk and investors will lock down the deal terms, with some eyeing computing power as part of their investment return. This power is key for firms creating new AI products.

French start-up Mistral looks set to challenge AI frontrunners Google and OpenAI

In Davos, AI was a hot topic among leaders, sparking debates about its future and necessary regulation. A survey revealed that CEOs expect AI to cut at least 5% of jobs this year. However, the buzz was about French startup Mistral, challenging the AI dominance of Google and OpenAI. Mistral's showing promise despite being new and facing competition from other startups like Toronto's Cohere. Even celebrities like will.i.am chimed in, seeing AI as a barrier-breaker. Innovations range from AI matching top students in geometry to potential animal language translation. Yet, there are concerns about AI perpetuating biases.

OpenAI CEO Sam Altman is still chasing billions to build AI chips

Sam Altman, CEO of OpenAI, is hustling to gather billions for making AI chips. He plans to set up a global network of factories with top chipmakers to meet the high demand for AI computations, like those used by ChatGPT and DALL-E. AI's heavy reliance on chips, especially Nvidia's H100 GPUs, has spiked Nvidia's value. The limited chip-making facilities have everyone, including Altman, scrambling to secure production capacity. OpenAI, needing big bucks, is in talks with big investors like SoftBank. Meanwhile, tech giants like Microsoft, Amazon, and Google are also crafting their own AI chips, heating up the competition. Nvidia stays ahead, announcing advanced chips, while rivals AMD, Qualcomm, and Intel are also jumping into the fray with their AI processors.

AI startup Cohere in talks to raise roughly $500mln to $1bln

Cohere, an AI startup based in Toronto, is in talks to raise between $500 million and $1 billion. This move follows a surge in AI investments, partly due to the success of ChatGPT. In a previous funding round, Cohere raised $270 million, valuing it at $2.2 billion. 

The company, competing with firms like OpenAI, focuses on AI applications for businesses to enhance efficiency. Cohere also aims to use the funds for AI development, hiring talent, and expanding sales. Their technology is offered across various cloud platforms, and they've avoided financing through cloud usage credits.


The paper "LEGO: Language Enhanced Multi-modal Grounding Model" proposes LEGO, a new multi-modal model. Unlike existing models that focus on global information in each modality, LEGO excels at understanding detailed local information. It's adept at identifying and localizing specific regions in images or moments in videos. This is achieved through a unique dataset construction pipeline, creating a diverse multi-modal and multi-granularity dataset. This advancement allows for more precise multi-modal understanding. 

The paper "Jump Cut Smoothing for Talking Heads" introduces a novel method for enhancing the viewing experience of talking head videos by smoothing jump cuts. This technique involves fusing mid-level representations derived from DensePose keypoints and face landmarks with the appearance of the subject from other frames in the video. Key points and landmarks are interpolated for motion, and an image translation network synthesizes pixels. A unique cross-modal attention scheme corrects keypoint errors, choosing the best source for each key point, resulting in more seamless transitions even in challenging scenarios like drastic movements or rotations of the head. 

The paper discusses enhancing Large Language Models' (LLMs) capacity to handle long inputs. It proposes a method to extend LLMs' context window beyond their pre-trained limits by modifying rotary position embedding (RoPE). The approach involves adjusting RoPE's base frequency and scaling attention logits, leading to more efficient adaptation to larger context windows. This method, tested on LLaMA-2-7B-Chat, extended its context to 16,384 using just 100 samples and 6 training steps, showing remarkable efficiency. The paper also explores the impact of data composition and training curriculum on this extension process.


Hermeso - AI-driven travel guide app that personalizes exploration experiences based on user preferences. It suggests places matching individual needs, including dietary restrictions and budget considerations, ensuring tailored discovery. 

Marblism - a platform that quickly launches React and Node.js apps. Users describe their app idea, and Marblism automates codebase creation, including database schema, API endpoints, and front-end pages, using AI and predefined templates.

Pinokio - a unique browser that enables users to install, run, and control any application with just one click. It simplifies the process of managing applications by automating installation and operation, catering to users who seek ease and efficiency in computer application management.

Locofy - AI tool that significantly accelerates frontend development by converting designs to code in one click. It integrates with design tools like Adobe XD and Figma, supports various technologies, and is powered by LocoAI. 

Sama AI - AI-driven personal companion app designed to enhance memory and self-improvement. It captures conversations, generates insights, and helps users connect with past experiences. 



A New Nonprofit Is Seeking to Solve the AI Copyright Problem

Ed Newton-Rex quit his AI executive role over copyright disagreements. He objected to AI models being trained on content scraped from the internet without creators' consent. Lawsuits against major AI firms have emerged, challenging this practice. Newton-Rex founded ‘Fairly Trained,’ a nonprofit certifying AI companies using consensually sourced data. This move aims to promote ethical data use and support creators' rights. The initiative highlights the ongoing debate about fair use and creator compensation in the AI industry. TIME

Saudi Arabia goes big in Davos as it looks to become a top AI tech hub for the Middle East

Saudi Arabia showcased its ambitions to become a top AI tech hub in the Middle East at the World Economic Forum in Davos. Highlighting projects like Neom, it seeks to diversify its economy beyond oil, focusing on technology and innovation. This initiative challenges the UAE's current tech dominance in the region. Saudi Arabia's efforts include massive infrastructure projects, attracting fintech talent, and promoting education and entrepreneurship, particularly in AI, to realize these goals by 2030. CNBC

AI Unlocks Secrets of Human Imagination and Memory Formation

A UCL study uses AI to understand human memory and imagination. By simulating the hippocampus and neocortex, researchers show how the brain forms and retrieves memories, enabling both recollection of past events and creation of new scenarios. This research offers insights into memory's role in survival and predicting future events, and how memory distortions occur. The study, advancing our understanding of memory processing, has significant implications for neuroscience. NEUROSCIENCE NEWS

How satellite images and AI could help fight spatial apartheid in South Africa  

In South Africa, AI and satellite imagery are being used to combat spatial apartheid. Raesetje Sefala and her team at DAIR analyze satellite images to reveal the unequal distribution of land and resources between townships and wealthier suburbs. Their goal is to provide data that can influence policy and improve resource allocation, addressing the enduring inequalities stemming from apartheid. This innovative approach seeks to make visible the overlooked needs of marginalized communities. MIT

Using generative AI to identify potent and selective MYT1 inhibitors for the treatment of cancer

Insilico Medicine's AI-driven research identified MYT1 as a key target for treating breast and gynecological cancers. Their AI platforms, PandaOmics and Chemistry42, facilitated the discovery of novel, selective MYT1 inhibitors. These inhibitors, particularly Compound 21, show potential in addressing cancers with genome instability, like CCNE1 amplification. Compound 21 demonstrated good MYT1 activity, high selectivity, and promising preclinical efficacy, paving the way for safer, more effective cancer treatments. This approach exemplifies the power of AI in medicinal chemistry and cancer therapy development. PHYS ORG

Both of these agencies want a piece of Microsoft’s Open AI partnership

The Justice Department and the Federal Trade Commission are in discussions to determine which agency will investigate the partnership between Microsoft and OpenAI, including ChatGPT, on antitrust grounds. Both agencies are interested in probing the collaboration, which includes Microsoft's significant investment in OpenAI and their shared technological developments. Concerns revolve around potential unfair market advantages in AI technology. The situation is further complicated by Microsoft's involvement in the rehiring of OpenAI CEO Sam Altman. This inquiry is part of a broader focus on competition in the AI sector by both agencies. POLITICO

    What'd you think of today's edition?

    Login or Subscribe to participate in polls.

    What are MOST interested in learning about AI?

    What stories or resources will be most interesting for you to hear about?

    Login or Subscribe to participate in polls.


    or to participate.