Explore

Slice the corpus on any axis.

Filter by firm, year, function, metric or slide category — then see the result as a narrative arc, a treemap of slide types, or a catalog of slides.

clear
51,416 slides across 1229 decks match.
Treemap Arc Catalog
But how 'open' are 'open source' models?
AirStreetCapital · comparison_table
But how 'open' are 'open source' models?
 Is contamination inflating progress?
AirStreetCapital · diagnosis
Is contamination inflating progress?
 Researchers try to correct problems in widely used benchmarks
AirStreetCapital · problem_statement
Researchers try to correct problems in widely used benchmarks
57% — error rate
 Live by the vibes, die by the vibes...or close your eyes for a year and OpenAI is still #1
AirStreetCapital · industry_trends
Live by the vibes, die by the vibes...or close your eyes for a year and OpenAI is still #1
 Are neuro-symbolic systems making a comeback?
AirStreetCapital · case_study
Are neuro-symbolic systems making a comeback?
25 — Olympiad-level geometry problems solved
 It’s possible to shrink models with minimal impact on performance...
AirStreetCapital · industry_trends
It’s possible to shrink models with minimal impact on performance...
40x — Model performance
 ...as distilled models become more fashionable
AirStreetCapital · industry_trends
...as distilled models become more fashionable
 Models built for mobile compete with their larger peers
AirStreetCapital · industry_trends
Models built for mobile compete with their larger peers
 Strong results in quantization point to an on-device future
AirStreetCapital · industry_trends
Strong results in quantization point to an on-device future
 Will representation fine tuning unlock on-device personalization?
AirStreetCapital · diagnosis
Will representation fine tuning unlock on-device personalization?
 Hybrid models begin to gain traction
AirStreetCapital · industry_trends
Hybrid models begin to gain traction
 And could we distill transformers into hybrid models? It's...complicated.
AirStreetCapital · case_study
And could we distill transformers into hybrid models? It's...complicated.
3B — Average Accuracy (%)
 Either way, the transformer continues to reign supreme (for now)
AirStreetCapital · industry_trends
Either way, the transformer continues to reign supreme (for now)
74% — Research paradigm share
 Synthetic data starts gaining more widespread adoption...
AirStreetCapital · industry_trends
Synthetic data starts gaining more widespread adoption...
 ...but Team Model Collapse isn’t going down without a fight
AirStreetCapital · context
...but Team Model Collapse isn’t going down without a fight
 Web data is decanted openly at scale - proving quality is key
AirStreetCapital · case_study
Web data is decanted openly at scale - proving quality is key
 Retrieval and embeddings hit the center stage
AirStreetCapital · industry_trends
Retrieval and embeddings hit the center stage
 Context proves a crucial driver of performance
AirStreetCapital · diagnosis
Context proves a crucial driver of performance
35% — retrieval failure rate
 Evaluation for RAG remains unsolved
AirStreetCapital · problem_statement
Evaluation for RAG remains unsolved
 Frontier labs face up to the realities of the power grid and work on mitigations
AirStreetCapital · industry_trends
Frontier labs face up to the realities of the power grid and work on mitigations
 Could better data curation methods reduce training compute requirements?
AirStreetCapital · diagnosis
Could better data curation methods reduce training compute requirements?
23 — FLOPs %
 Chinese (V)LLMs storm the leaderboards despite sanctions
AirStreetCapital · industry_trends
Chinese (V)LLMs storm the leaderboards despite sanctions
 And Chinese open source projects win fans around the world
AirStreetCapital · industry_trends
And Chinese open source projects win fans around the world
 VLMs achieve SOTA performance out-of-the-box
AirStreetCapital · case_study
VLMs achieve SOTA performance out-of-the-box
 Diffusion models for image generation become more and more sophisticated
AirStreetCapital · case_study
Diffusion models for image generation become more and more sophisticated
 Stable Video Diffusion marks a step forward for high-quality video generation...
AirStreetCapital · case_study
Stable Video Diffusion marks a step forward for high-quality video generation...
 ...leading the big labs to release their own gated text-to-video efforts
AirStreetCapital · industry_trends
...leading the big labs to release their own gated text-to-video efforts
 Meta goes even further, throwing audio into the mix
AirStreetCapital · case_study
Meta goes even further, throwing audio into the mix
 AI gets en-Nobel-ed
AirStreetCapital · case_study
AI gets en-Nobel-ed
 AlphaFold 3: going beyond proteins and their interactions with other biomolecules
AirStreetCapital · case_study
AlphaFold 3: going beyond proteins and their interactions with other biomolecules
 ...starting a race to become the first to reproduce a fully functioning AlphaFold3 clone
AirStreetCapital · industry_trends
...starting a race to become the first to reproduce a fully functioning AlphaFold3 clone
 AlphaProteo: DeepMind flexes new experimental biology capabilities
AirStreetCapital · case_study
AlphaProteo: DeepMind flexes new experimental biology capabilities
300-fold — Experimental success rate
 The Bitter Lesson: Equivariance is dead...long live equivariance!
AirStreetCapital · industry_trends
The Bitter Lesson: Equivariance is dead...long live equivariance!
 Scaling frontier models of biology: EvolutionaryScale' ESM3
AirStreetCapital · case_study
Scaling frontier models of biology: EvolutionaryScale' ESM3
 Language models that learn to design human genome editors
AirStreetCapital · case_study
Language models that learn to design human genome editors
71.7% — Max A:G editing (%)
 Yet, evals and benchmarking in BioML remains poor
AirStreetCapital · problem_statement
Yet, evals and benchmarking in BioML remains poor
 Expanding the protein function design space: challenging folds and soluble analogues
AirStreetCapital · case_study
Expanding the protein function design space: challenging folds and soluble analogues
 Foundation models for the mind: learning brain activity from fMRI
AirStreetCapital · case_study
Foundation models for the mind: learning brain activity from fMRI
6,700 — fMRI reconstruction accuracy
 Foundation models across the sciences: the atmosphere
AirStreetCapital · case_study
Foundation models across the sciences: the atmosphere
5,000x — Computational speed
 Foundation models for the mind: reconstructing what you see
AirStreetCapital · case_study
Foundation models for the mind: reconstructing what you see
 Speaking what you think
AirStreetCapital · case_study
Speaking what you think
99.6% — Word Error Rate
 A new challenge aims to refocus the industry on the path to AGI
AirStreetCapital · case_study
A new challenge aims to refocus the industry on the path to AGI
46 — ARC-AGI benchmark score
 LLMs still struggle with planning and simulation tasks
AirStreetCapital · diagnosis
LLMs still struggle with planning and simulation tasks
12% — Accuracy percentage
 Can LLMs learn to think before they speak?
AirStreetCapital · industry_trends
Can LLMs learn to think before they speak?
 Open-endedness gathers momentum as a promising research direction
AirStreetCapital · case_study
Open-endedness gathers momentum as a promising research direction
 But were implicit reasoning capabilities staring us in the face the whole time?
AirStreetCapital · industry_trends
But were implicit reasoning capabilities staring us in the face the whole time?
 Program search unlocks new discoveries in the mathematical sciences
AirStreetCapital · case_study
Program search unlocks new discoveries in the mathematical sciences
0.03% — Fraction of excess bins
 RL drives improvements in VLM performance...
AirStreetCapital · case_study
RL drives improvements in VLM performance...
62.7% — task success rate
Next page →