AI & Startups May 5 - May 12

Gen AI Product News

OpenAI ChatGPT search feature coming Monday

According to multiple sources, OpenAI is planning to announce a new search feature for ChatGPT on Monday, directly competing with Google Search and Perplexity.

The details:

• The search functionality would allow ChatGPT to access online sources to answer user queries, providing citations for the info used.
• Some versions of the feature may also incorporate relevant images alongside text responses, such as diagrams or illustrated instructions.
• OpenAI currently offers limited browsing capabilities to ChatGPT Plus subscribers, which has been inconsistent and buggy.

SoundHound AI and Perplexity team up on voice AI

SoundHound AI just announced a new partnership with Perplexity, aiming to integrate the company’s online LLM capabilities with real-time web knowledge across voice assistants in cars and devices.

The details:

• SoundHound’s Chat AI is already integrated with several in-car assistants, restaurant ordering tech, and mobile phones, apps, and TVs.
• Perplexity’s answer-engine AI will enable SoundHound’s assistant to provide more accurate responses and handle more complex questions.
• Perplexity’s web access also allows for real-time information, syncing info like 'gas price trends’ with an in-car assistant to navigate to the nearest station.

Google DeepMind reveals AlphaFold 3

Google DeepMind and Isomorphic Labs just introduced AlphaFold 3, the newest version of the groundbreaking AI model that can predict the structure of proteins, DNA, and other molecules with extreme accuracy.

The details:

• AlphaFold 3 has a 50% improvement in predicting drug-like interactions compared to traditional methods.
• While AlphaFold 2 focused on protein structures, 3 can handle ‘all of life’s molecules’ and can model and predict complex interactions.
• The model is available freely for non-commercial use through the new AlphaFold Server, allowing scientists to generate predictions and accelerate research.
• Isomorphic Labs, a sister company of Google DeepMind, is already using AlphaFold 3 with pharmaceutical partners to design new drugs.

Apple unveils iPad AI upgrades

Apple just revealed its new line of iPads at a company event in Cupertino, CA — featuring a custom M4 chip that enables advanced AI capabilities and a slew of new AI-powered features.

The details:

• The bigger iPad Pro now features the M4 chip with an upgraded Neural Engine, which CEO Tim Cook calls “an outrageously powerful chip for AI”.
• The M4 is capable of handling 38T operations per second, 4x the performance of previous models, allowing for the running of advanced AI.
• New AI features on the Pro include a True Tone Flash for document scanning and new video, image, and music editing tools.
• Prior to the event, a new report revealed that Apple is developing its own AI chips for data centers under the code name "Project ACDC".

Microsoft training massive in-house LLM

Microsoft is reportedly developing a massive 500B parameter in-house LLM called MAI-1, aiming to compete with top AI models from OpenAI, Anthropic, and Google.

The details:

• MAI-1 is being led by ex-Google DeepMind founder Mustafa Suleyman, who recently joined the company following the acquisition of AI startup Inflection.
• MAI-1 is not based on Inflection’s models, though tech and training from the company may carry over to the new system.
• The 500B parameter model is larger than Microsoft's previous open-source models, with Phi-3 mini trained on just under 4B parameters.

OpenAI partners with Stack Overflow

OpenAI just announced a new partnership with the developer platform Stack Overflow to enhance the performance of its AI models in programming tasks.

The details:

• OpenAI will use Stack Overflow's OverflowAPI to access the platform's vast repository of technical Q&A data accumulated over 15 years.
• The partnership aims to make ChatGPT better at providing coding answers by leveraging Stack Overflow's community-verified content.
• Each ChatGPT response using Stack Overflow data will link back to relevant posts, fostering deeper engagement with the developer community.
• The move marks a major change for the platform, which previously banned AI responses —?though has since also partnered with Google’s Gemini.

Setu unveils Sesame, India's first domain-specific LLM for BFSI sector

Setu, a leader in the Indian fintech space, has launched Sesame, the first Large Language Model (LLM) designed specifically for the Banking, Financial Services, and Insurance (BFSI) sector, in collaboration with Sarvam AI.

The Details:

• Collaboration and Innovation: Developed with Sarvam AI, Sesame aims to revolutionize financial services, enhancing processes like credit underwriting and fraud detection through AI.
• Event and Endorsements: The unveiling occurred at the Adbhut India event, attended by key figures such as Nandan Nilekani of EkStep Foundation.
• Technology and Impact: Leveraging India's robust digital infrastructure, Sesame offers personalized, efficient financial services across various customer interactions.
• Strategic Goals: This initiative reflects Setu's commitment to democratize financial services through advanced, compliant AI solutions, aligning with India’s digital innovations like Aadhaar and UPI.
• Future Outlook: The successful implementation of Sesame could serve as a model for integrating AI across different sectors, utilizing India's digital public goods.

This launch marks a significant step in applying domain-specific AI solutions to India’s BFSI sector, setting new standards for personalized and data-driven financial services.

India's GenAI Chatbot 'Hanooman' Goes Live in 98 Languages

Hanooman, a new generative AI (GenAI) chatbot, has launched in 98 languages, including 12 Indian languages like Marathi, Gujarati, Bengali, and Kannada. Developed by Telangana-based SML India and Abu Dhabi's 3AI Holding, the chatbot is now available on the Google Play Store and will soon be accessible on iOS.

Key Highlights:

• Extensive Language Support: Available in a wide range of languages, Hanooman aims to democratize AI accessibility across diverse linguistic backgrounds.
• Strategic Partnerships: Collaboration with Hiranandani Group-backed Yotta leverages advanced GPU cloud infrastructure to enhance Hanooman's capabilities.
• Specialized Features: From casual chats to professional advice, and coding to tutoring, Hanooman is equipped to handle a variety of tasks. It also plans to introduce multimodal functionality, allowing it to process speech, audio, video, and images.
• Inclusivity and Expansion: With a vision to reach 200 million users within its first year, Hanooman focuses on making GenAI technology inclusive and widely available.
• Open-Source Development: Future plans include offering open-source LLMs to enterprises, targeting sectors like healthcare, governance, financial services, and education.

This launch represents a significant step in harnessing AI for multilingual and multicultural integration, positioning Hanooman as a pivotal player in the expanding Indian AI market.

Gen AI VC

UK AI startup Wayve raises $1.05B from SoftBank Group Corp. , 英伟达 , 微软 to accelerate autonomous mobility

Peak XV Partners -backed data & AI startup Atlan bags $105 million led by GIC, Meritech Capital

AI chip startup DeepX secures $80M Series C at a $529M valuation

Rad AI , a startup that helps radiologists save time on report generation, raises $50M Series B from Khosla Ventures

Israeli startup Panax raises a $10M Series A for its AI-driven cash flow management platform

Legaltech startup Definely has raised a $7m Series A to develop its AI-powered software that aims to help lawyers understand and edit documents

AI startup Myelin Foundry secures $4 mn in funding for Edge AI

ETH spinout Mimic raises $2.5m for AI-powered robotic hands

Mergers and acquisitions (M&A)

Samsung Research Medison to acquire French AI ultrasound startup SONIO for $92.7M

DocuSign acquires AI-powered contract management firm Lexion

AI RESEARCH

Meta Llama-3 goes head-to-head with OpenAI GPT-4

Lmsys just published a new deep dive into its Chatbot Arena data, comparing Meta’s open-source Llama 3 70B model against top models like GPT-4 — revealing surprising strengths in the open-source leader.

The details:

• Llama-3 is the top open-source model on the Lmsys Leaderboard, featuring over 70,000 votes from users in Arena Battle testing.
• Meta’s model shines in battles involving brainstorming and writing prompts, falling short in math, coding, and translation compared to top competitors.
• Lmsys noted Llama-3’s tone is friendlier and more conversational than rivals, which the model exhibits in its victorious battles in the Arena.

Teachers can’t detect AI writing

A series of studies from several German universities found that both novice and experienced teachers struggled to accurately distinguish between student-written and AI-generated texts.

The details:

• The first study trialed pre-service teachers on AI and student text, correctly ID’ing only 45.1% of AI-generated submissions.
• The second study tested experienced teachers, who correctly identified just 37.8% of AI-generated submissions.
• The study revealed that teachers were overconfident in their ability to spot AI-generated text.
• AI detection software also often falsely flags student-written texts as potentially AI-generated, which further complicates the issue.

LLM agent writes code to train robots

Nvidia researchers just introduced DrEureka, an AI system that uses LLM agents to automate the process of training robot skills in simulations and transferring them to the real world.

The details:

• DrEureka writes code to teach robot skills in simulations, then ensures skills work well when transferred to physical robots.
• The system trained a simulated robot dog to balance and walk on a yoga ball, with the task then transferred successfully to a real robot in testing.
• DrEureka uses LLMs’ physics knowledge to adjust parameters like friction and gravity, automating the time-consuming process of adapting to the real world.

AI Adoption

AI usage surges in the workplace

Microsoft and LinkedIn just published their Work Trend Index Annual Report, revealing that AI adoption is surging in the workplace — calling 2024 the ‘year AI at work gets real’.

The details:

• The report found that use of GenAI has doubled in the last six months, with 75% of knowledge workers using the tech in some capacity.
• 78% of AI users are bringing their own AI to work — with 52% reporting they are reluctant to admit to its use.
• 66% of leaders wouldn't hire someone without AI skills, and 71% prefer less experienced candidates with AI aptitude over a more experienced one without it.
• AI power users reported enhanced productivity, creativity, and job satisfaction compared to skeptical peers.

How organizations have benefited from generative AI adoption

More than six in 10 organizations who use AI say it’s improved efficiency, and 59% say it’s improved customer experience, according to a new report from Scale AI. In addition, nearly half of respondents say LLMs have helped them come up with ideas for new products or services — a testament to AI’s brainstorming capabilities.

Only around a third of companies say their profit and revenue have improved thanks to AI. That shows there’s still room for improvement when it comes to AI’s money-making potential. Still, just 8% of organizations said they haven’t seen any improvements across the board. Seven percent have yet to adopt AI in any capacity.

Companies are starting to adopt different types of AI models as well. In 2023, only 37% of companies used commercial or closed-source AI platforms. But by 2024, that figure rose by nearly 50 points, with roughly 86% of companies relying at least in part on closed-source LLMs.

At the same time, open-source models have gotten more popular, too. Today, two-thirds of organizations use open platforms, compared with just 41% last year.

There might be opportunities for growth when it comes to AI platforms that have been built from the ground up to perform specific tasks. Around a quarter of organizations are training their own AI models, only a modest increase from 22% last year.

Words associated with AI are becoming more common in academic writing

Vocabulary commonly used in AI chatbots like ChatGPT is showing up more frequently in newly published research, according to a recent analysis by University College London librarian Andrew Gray.

Gray found that certain words associated with AI, including “intricate,” “meticulously,” and “commendable” became much more common in academic writing in 2023 compared with 2022.

That suggests that researchers are increasingly turning to ChatGPT for help writing their papers, raising important questions about the reliability of scientific writing in the age of AI.

Top AI research areas

Between 2017 and 2022, AI research focused most heavily on large language models like OpenAI’s GPT-4, with around 13,300 papers exploring the subject. Roughly 10,300 articles focused on distributed machine learning, when a model is trained across multiple computers.

Neural scene representation (a method of building 3D objects out of individual points) and plant disease detection were the third and fourth most popular research areas respectively.