New Technology / Ai Development
Track AI development, model progress, product releases, infrastructure shifts and strategic technology signals across the artificial intelligence sector.
Approaching the AI Event Horizon? Part 2, w/ Abhi Mahajan, Helen Toner, Jeremie Harris, @8teAPi
Topic
AI Advancements and Challenges
Key insights
- Abhi Mahajan develops foundation models at Noetic AI to predict cancer treatment responses, reflecting both skepticism and optimism about AIs impact on biology
- Helen Toner highlights CSETs report on automated AI R&D, emphasizing its role as a source of strategic surprise due to the lack of consensus on its effects
- Jeremie Harris discusses the challenges of controlling superhuman AI systems and the need for improved US-China collaboration on AI issues
- The episode reveals significant disagreement among AI experts about future developments, complicating situational awareness in a rapidly evolving field
- Nathan introduces large language models to identify blind spots in AI research, recommending the blind spot finder feature in GrenoLa for insights
- Abhi Mahajans work includes a competitive intelligence platform to enhance clinical analysis pipelines for identifying promising cancer drugs
Perspectives
short
Proponents of AI Development
- Highlight the transformative potential of AI in various fields
- Emphasize the importance of continual learning in AI for personalized user experiences
- Argue that market acceptance is driving rapid advancements in AI technology
- Point out the effectiveness of export controls in slowing down adversarial AI development
- Stress the need for proactive measures to ensure security in AI infrastructure
Skeptics of AI Progress
- Question the reliability of AI predictions and the validity of current models
- Express concerns about the complexities of biological systems and the limitations of AI in drug discovery
- Warn against the potential for misinterpretation of AI-generated insights
- Critique the assumption that user acceptance will lead to successful AI adoption
- Raise doubts about the effectiveness of treaties in managing international AI competition
Neutral / Shared
- Acknowledge the ongoing debate about the pace and impact of AI advancements
- Recognize the challenges in aligning research with market demands
- Note the importance of understanding the infrastructure that supports AI development
Metrics
other
a major source of potential strategic surprise
Helen Toner's report on automated AI R&D
This indicates the unpredictability of AI advancements and their implications.
automated AI R&D is simply a major source of potential strategic surprise.
other
the blind spot finder feature in GrenoLa
Nathan's recommendation for identifying blind spots in AI research
This tool could enhance situational awareness in a rapidly evolving field.
I am really enjoying the blind spot finder recipe that I recently created on GrenoLa.
other
a competitive intelligence platform to enhance clinical analysis pipelines
Abhi Mahajan's work at Noetic AI
This could significantly improve the identification of promising cancer drugs.
you built an entire competitive intelligence platform, Ellen based, to feed the clinical analysis pipeline.
failure_rate
97%
percentage of oncology trials that fail
This highlights the urgent need for improved methodologies in drug development.
97% of oncology trials fail.
other
lower cost of human labor
cost advantages in clinical trials
This cost reduction can lead to more trials being conducted.
partially because of like lower cost of human labor
other
more hands
workforce availability in China
A larger workforce can accelerate drug development processes.
they just have more hands
other
super long feedback loop
comparison of drug design processes
Long feedback loops can hinder innovation in drug development.
you have a super long feedback loop
automation
over 50%
cut help desk tickets
This indicates a significant reduction in workload for IT teams.
cut help desk tickets by more than 50%
Key entities
Timeline highlights
00:00–05:00
Abhi Mahajan is developing foundation models at Noetic AI to enhance predictions of cancer treatment responses. The discussions highlight the transformative potential of AI in biology and medicine, despite existing skepticism among experts.
- Abhi Mahajan develops foundation models at Noetic AI to predict cancer treatment responses, reflecting both skepticism and optimism about AIs impact on biology
- Helen Toner highlights CSETs report on automated AI R&D, emphasizing its role as a source of strategic surprise due to the lack of consensus on its effects
- Jeremie Harris discusses the challenges of controlling superhuman AI systems and the need for improved US-China collaboration on AI issues
- The episode reveals significant disagreement among AI experts about future developments, complicating situational awareness in a rapidly evolving field
- Nathan introduces large language models to identify blind spots in AI research, recommending the blind spot finder feature in GrenoLa for insights
- Abhi Mahajans work includes a competitive intelligence platform to enhance clinical analysis pipelines for identifying promising cancer drugs
05:00–10:00
Drug discovery processes are often inefficient, relying heavily on personal networks and clinical trial platforms. The complexity of identifying effective drugs is compounded by economic factors and the inherent challenges of biological research.
- Drug discovery is inefficient, often relying on personal networks. Scraping the semantic web for drug annotations could enhance efficiency
- AI models for drug recommendations are imperfect and require human oversight. Their evaluation is often inconsistent and subjective
- Identifying effective drugs is complex, influenced by company economics and relationships. This complicates the drug acquisition process
- Interest in autonomous AI systems for experiments is growing, but biologys lack of verifiable ground truth hinders progress
- Reliable data in biology takes time to gather, slowing AI learning. It can take 18 months to obtain a single actionable data point
- Toxicology presents challenges for AI due to drug variability. In vivo observations are often necessary to understand toxicity
10:00–15:00
Isomorphic Labs has developed a predictive model that reportedly doubles AlphaFold's performance, which raises questions about its practical utility in drug development. Despite advancements, the primary challenge remains the translation of pre-clinical assets into effective patient outcomes.
- Isomorphic Labs predictive model reportedly doubles AlphaFolds performance, raising questions about its utility in drug development
- Despite advancements, the bottleneck remains in translating pre-clinical assets to improved patient outcomes
- Optimizing every aspect of proteins in the pre-clinical pipeline could significantly benefit drug development
- Major biological advancements often arise from improved imaging and sensing techniques, enhancing understanding of biological processes
- Future models may simulate human in vivo biology to address the high failure rate of oncology trials
- Collecting data from real human tumors could lead to a genuine human simulator, enhancing research validity
15:00–20:00
97% of oncology trials fail, indicating a significant challenge in drug design and patient response prediction. NOEDEC aims to collect extensive human tumor data to enhance understanding of patient responses.
- 97% of oncology trials fail, highlighting the need for improved drug design and patient response prediction. NOEDEC aims to collect extensive human tumor data to enhance understanding of patient responses
20:00–25:00
NOEDEC utilizes tumor profiling to predict cancer treatment responses by identifying unique biomarkers. This approach may enhance the understanding of patient responses, although the interpretability of the model remains a concern.
- NOEDEC predicts cancer treatment responses by profiling tumor samples, identifying unique biomarkers that may not be human-interpretable. This could enhance patient response prediction
25:00–30:00
Fine-tuning language models at test time can optimize answers for specific problems, potentially enhancing predictions for cancer patients. The effectiveness of this approach remains to be validated in practical applications.
- Fine-tuning language models at test time can optimize answers for specific problems, potentially enhancing predictions for cancer patients