Beginner Lessons

AI Trends 2024: Machine Learning & Deep Learning with Thomas Dietterich - 666

LangChain vs LlamaIndex. Beyond data, constructing context & post-filtering enhance accuracy from VB

Real-world AI demands live data, not static data corpus. Fine-tuning sensitive data leads to leakage

Natural program integrates deductive verification into the chain of thought, self-verified by LLM.

AI Trends 2024: Computer Vision with Naila Murray - 665

Prediction of shift in LLMs: Embedding desiderata in training to avoid post hoc interventions.

The greatest leverage towards generative AI challenges means defining specific use cases and metrics

Markus' approach: not quantization method but a solution to eliminate activation quantization issues

The power of generative models is open-endedness, but this also poses challenges for responsible AI

Success lies in listening, engaging, adapting to evolving needs and building sticky, useful products

Responsible AI in the Generative Era with Michael Kearns - 662

Making learning enjoyable and using play as motivation sparks increased engagement and interest

PartyRock sparks generative AI intuition through playful, no-code app building—unleashing creativity

Coactive transforms workflows, shifting from TLS to LST, revolutionizing image & video data handling

Building at Coactive is like building a car: Model-agnostic for easy upgrades and future-proof ready

Data, Systems and ML for Visual Understanding with Cody Coleman - 660

Explore diverse tool collaboration for enhanced LLM solutions, usability, and dev-friendly solutions

Retrieval patterns allows the model to act as a ‘problem-solving brain’, yielding smarter solutions.

AI should ease burdens on users. Diverse data is key for fairness, inclusivity, and accessibility.

Prem highlights “invention rich areas” in AI and invention challenges as these technologies evolve

Boost AI by working with academia to define targeted datasets and foster open synthetic data culture

Ensure startup relevance by innovating beyond one feature and aiming for long-term value and impact

Engineers use simulation to tackle bias and adversarial examples, and enhance robust model solutions

Applying MLOps techniques enhances speed and consistency in deploying edge and embedded AI systems

Building a unicorn team: ✅ Diverse skills ✅ Open communication ✅ Continuous learning

The AI moratorium letter signals existential risks to the public and urges caution in AI development

Yoshua calls for enhanced regulations, investments, and countermeasures for AI safety and defense

AI concerns demand traceability. Resource allocation sparks values divide in present & future models

The main challenges for multilingual models: balancing languages and dealing with poor data quality.

Turn problems to checklists and restructure from low to high-level judgments to avoid hallucinations

Anthropomorphism in terminology is inevitable but it’s important to clarify these are just metaphors

Prompting pretrained models is subtractive, unveiling the model's behavior and text predictions

Finding global talent and open collaboration are keys to establishing unique and meaningful research

Evaluating AI's general performance & finding models for unrestricted prompting remains a challenge

AI shifts from engineering to complex system science, struggles with understanding emergent behavior

Shift to multimodal models: Visual grounding, embodiment, & more data unlock exciting possibilities

Data transparency fosters accountability. Awareness & change in scientific practices drive fairness

Reimagining AI: A tool for empowerment, not inevitability

Turning researchers into fighters and fighters into researchers and building technology for people

Altering the weights is vital in adding concepts. Hypernetwork personalize weights for faster 🖼️ gen

Through fine-tuning and specialization, you resurface the prior knowledge leading to better results.

The core idea behind diffusion models: Clean ➡️ Noise ➡️ Train ➡️ Denoise ➡️ Generate

The ML metrics are behind where the tech is. LLMs score lower on metrics but generate better results

"LLMs are narcissistic in that they find their own work better than a human-generated text” - Shreya

Is RAG an essential component of the LLM stack? #llm #aisafety #artificialintelligence

Hallucination is a top concern in LLM safety but broader AI safety issues lie beyond hallucinations

Language 🗣️ is the key ingredient to achieving human-like AI 🧠

Specific and meaningful labels instead of generic labels leads to a richer language representation