In this week in AI research, OpenAI’s latest models impress in some STEM related tasks, especially in coding. Math is another strong point. In addition, Salesforce is making good on its promise to base its company on AI “agents” – autonomous entities handling customer service and scheduling. CEO Marc Benioff posits them as a scalable alternative to human hiring. Meanwhile, LinkedIn has revealed plans to tap user data in forthcoming AI services. Privacy concerns ensue. In biopharma, Insilico Medicine reported positive Phase IIa results for its AI-designed therapy targeting idiopathic pulmonary fibrosis, a condition with limited treatment options and poor outcomes.

Why Salesforce’s AI strategy could translate to fewer humans, more efficiency for businesses

Source: Bloomberg (September 17, 2024)

At its Dreamforce conference, Salesforce Inc. formalized a shift in its strategy that it had hinted at before: making AI a foundation of its future offerings rather than just a feature. Specifically, the SaaS company unveiled a suite of generative AI “agents” that can independently perform tasks previously handled by human employees. Those include resolving customer service inquiries to scheduling sales meetings. This move, according to CEO Marc Benioff, empowers businesses to achieve unprecedented scalability, navigating peak demand periods without resorting to traditional hiring practices. It will also lead to reduced need, Benioff said, billing it as a point of differentiation. While prior iterations of AI assistants, including Microsoft’s Copilot, relied on human intervention, Salesforce’s latest AI products aim to operate more independently and deliver tangible results — namely, more efficiency and reduced need for human labor. For Salesforce, too, the strategy aims to safeguard its revenue even if AI leads to reduced software subscriptions as a result of workforce reductions. Other tech companies such as Meta have envisioned a world in which AI agents proliferate, potentially even outnumbering humans.

YouTube announces new genAI features for video, music, and inspiration

Source: NBC News (September 18, 2024)

At a promotional event, YouTube CEO Neal Mohan unveiled a suite of generative AI features, including text-to-video capabilities for Shorts, AI-powered content suggestions, and enhanced auto-dubbing with more natural and expressive speech. These tools, built on Google’s DeepMind technology, aim to empower creators by boosting productivity and creativity. To address concerns about AI-generated disinformation, YouTube will implement SynthID watermarks on AI-created content. In addition, YouTube is adding new monetization options for creators. The Alphabet subsidiary is also introducing new community engagement features.

Insilico Medicine hits Phase IIa milestone with AI-designed IPF drug

Source: Insilico Medicine press release (September 18, 2024)
Insilico Medicine recently announced positive preliminary results from a Phase IIa clinical trial of ISM001-055, its AI-designed therapy for idiopathic pulmonary fibrosis (IPF). The 12-week study demonstrated that ISM001-055 was safe and well-tolerated across all doses. The drug also showed a dose-dependent improvement in lung function, with the highest dose (60 mg once daily) yielding the most significant improvements. Insilico plans to consult with regulatory authorities to design a Phase IIb study. In related news, Alex Zhavoronkov, founder and CEO of Insilico Medicine, pointed out that the company’s 2019 proof-of-concept published in Nature Biotechnology, titled “Deep learning enables rapid identification of potent DDR1 kinase inhibitors,” was initially met with skepticism from medicinal chemists. Since then, generative AI technology has gained traction and is emerging as a mainstream tool in drug molecule selection processes.

LinkedIn to tap user data for AI training

Source: Fast Company (September 19, 2024)

LinkedIn is changing its user agreement, beginning November 20, 2024, to automatically use personal data and content to train its generative AI models. This change applies to users in the U.S., Canada, and several other countries, but not the European Union or Switzerland where privacy regulations are stricter. The organization claims the move will improve its AI-powered products and help creators. Users can opt out by adjusting settings on both desktop and mobile. While LinkedIn’s Chief Privacy Officer assures users that privacy-enhancing technologies are in place to minimize personal data in these datasets, the plans have sparked privacy concerns among some users, as Fast Company noted.

ChatGPT o1-preview bests Anthropic’s Sonnet 3.5 in comparative AI Model Tests

Source: LinkedIn Post by Daniel Schauer (September 13, 2024)
In a recent comparison of leading closed-source AI models, OpenAI’s ChatGPT o1-preview emerged as the clear winner over Anthropic’s Sonnet 3.5. Daniel Schauer, Digital Rep CoPilot Platform Manager at Takeda, pitted the performance of the models against each other across various tasks, including coding, language understanding, and simulation creation. OpenAI’s latest model consistently demonstrated superior accuracy and user experience, excelling in generating well-structured code and handling complex prompts. Schauer noted that ChatGPT o1-preview was “particularly adept at answering detailed questions and producing polished outputs.” He concluded, “The clear winner is ChatGPT o1-preview.” While Sonnet 3.5 attempted to improve its responses through iterative testing, it struggled to maintain accuracy. OpenAI had noted that o1-preview scored 83% on the American Invitational Mathematics Examination (AIME), a USA Math Olympiad qualifier. A smaller reasoning model, o1-mini, scored 70%, placing it among the top 500 U.S. high school students. GPT-4o only managed 12%. On Codeforces, a competitive coding platform, o1-preview reached the 89th percentile. o1-mini achieved an Elo rating of 1650, placing it in the 86th percentile.

ChatGPT’s o1-preview model can crack some Ph.D.-level physics problems

Source: YouTube (September 13, 2024)
“One Jackson problem takes an average of one and a half weeks to finish,” said Kyle Kabasares, Ph.D., who gave OpenAI’s o1-preview model a trio of problems from the famously difficult physics graduate textbook “Classical Electrodynamics” from John David Jackson, Ph.D., a late Canadian–American physicist. In a test from the middle of the book, the o1-preview model solved the problem in 21 seconds. After entering the problems in a mix of text and LaTeX, Kabasares watched as the model reasoned through the problems step-by-step, ultimately arriving at the correct solutions in a fraction of the time it would take even the brightest graduate student — five minutes total for all three questions. Some viewers of the video expressed skepticism, questioning whether o1 is genuinely reasoning or simply regurgitating its training data to mimic problem-solving.

In a separate video, Kabasares challenges the o1-preview model to recreate the Python code he developed for his Ph.D. research on black hole mass measurements. After giving the model the entire method section from his LaTeX documentation, Kabasares sits back while watching the model generating lines of code with skepticism. “It’s probably going to fail,” he said. “There’s just so many things you need to know,” he added. “It took me a year to write this code. This was my baby. I published two papers with it.” Kabasares shares that his program had 1144 lines of code while the o1 model was closer to 100.

Billionaire Tech CEO Jim Kavanaugh urges transparency on AI’s impact on jobs

Source: CNBC (September 19, 2024)

World Wide Technology (WWT) CEO and billionaire Jim Kavanaugh stressed the importance of honesty and transparency regarding AI’s impact on the workforce in a recent CNBC interview. “If you think you’re going to tell employees nothing’s going to change, that’s just BS,” Kavanaugh stated. While acknowledging AI will disrupt certain roles, Kavanaugh remains optimistic about its potential to enhance productivity. “I truly believe it will be an enhancer and an accelerator,” he said. Kavanaugh encouraged employees and employers to become “students of AI and tech” to navigate the changes.

Asimov launches AI-powered AAV Edge platform for gene therapy design and manufacturing

Source: BioPharmaTrend.com (September 19, 2024)

The synthetic biology firm Asimov has unveiled AAV Edge, a platform designed to optimize the design and manufacturing of adeno-associated virus (AAV) gene therapies. AAV Edge tackles common challenges in gene therapy production, such as safety, manufacturability, and cost. A standout feature of AAV Edge is its AI-designed tissue-specific promoters (TSPs), which it has validated in preclinical studies to ensure targeted gene expression. These promoters have demonstrated more than a 200-fold dynamic range in expression between target and off-target tissues. Additionally, AAV Edge includes a DNA optimization tool that taps Asimov’s proprietary Kernel software, achieving up to a 7x increase in transgene expression in HEK293T cells.

CyberAIDD unveils AI platform to accelerate early-phase drug discovery

Source: LinkedIn (September 19, 2024)
CyberAIDD has launched an innovative platform that leverages advanced AI technologies to streamline early-phase drug design processes. Aimed at boosting research efficiency, the platform assists in identifying First-in-Class (FIC) drug targets and optimizing molecular structures. Key features include the CyberX-SAR and CyberX-Discovery tools, which utilize AI to enhance structure-activity relationships and lead optimization. The company positions its solution as a revolutionary approach to drug discovery, potentially expediting the development of new therapies.

New “StreamingLLM” technique enables infinite text input for language models

Source: LinkedIn Post by AlphaSignal (September 19, 2024)
A novel technique called “StreamingLLM” has been introduced, allowing large language models (LLMs) to handle infinite text input without any loss in accuracy. The method works by identifying key tokens that guide the model’s decisions and caching recent tokens, resulting in up to 22 times faster inference compared to traditional LLMs. This breakthrough could significantly enhance the efficiency of LLMs in processing long documents or continuous streams of text, marking a notable advancement in AI capabilities.

Moderna to cut R&D spending by 20%, refocuses pipeline on key products

Source: BioPharma Dive (September 18, 2024)
Moderna, a company that has made AI an educational priority for employees, is bowing to pressure from investors to cut its R&D spending by 20% and streamline its pipeline. The biotech intends to shift resources toward products that are nearing approval or already on the market, with a particular focus on oncology. Annual R&D expenses are projected to decrease from an estimated $4.8 billion in 2024 to between $3.6 billion and $3.8 billion by 2027. Moderna will discontinue five clinical programs but does not anticipate layoffs as part of the restructuring. The company also adjusted its revenue outlook, projecting $2.5 billion to $3.5 billion in sales for 2025 and aiming to break even by 2028.

French AI ecosystem doubles in size, raising $1.4 billion in funding

Source: LinkedIn Post shared by Yann LeCun (September 14, 2024)
The French generative AI (GenAI) ecosystem has seen significant growth over the past year, with the number of GenAI companies increasing from 73 in 2023 to 150 in 2024. Together, these companies have raised $1.4 billion in funding. The expansion includes foundational models beyond language, encompassing images, videos, biology, and chemistry. The top four largest GenAI fundraisings in France are centered on foundational models. This growth signifies a maturing ecosystem with more verticals being disrupted by AI applications.

Lionsgate partners with AI research company Runway prompting concerns in animation and VFX industries

Source: WSJ (September 18, 2024)

The studio Lions Gate Entertainment has signed a deal with AI startup Runway, granting the tech company access to its library of movies and shows in exchange for a custom-built AI model. The studio plans to use this new AI tool for internal purposes such as storyboarding in the creation of new movies and TV shows. This partnership signals the growing adoption of generative AI in Hollywood, potentially reshaping production processes in the film and animation industries. The move has raised concerns among animation and visual effects professionals about job security and the ethical implications of AI-driven content creation.

Research assistance: Frederic Célerse, Ph.D., Research Scientist in AI for Chemistry, Ecole polytechnique fédérale de Lausanne