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
The concerning phenomenon of AI psychosis
AirStreetCapital · problem_statement
The concerning phenomenon of AI psychosis
 The Model Welfare debate: what does the opposition think?
AirStreetCapital · context
The Model Welfare debate: what does the opposition think?
 Just Say No: Claude earns the right to end dangerous conversations
AirStreetCapital · case_study
Just Say No: Claude earns the right to end dangerous conversations
60.0% — % of Conversations Ended
 Single point of failure: how LLM safety mechanisms can be directly disabled
AirStreetCapital · diagnosis
Single point of failure: how LLM safety mechanisms can be directly disabled
99% — Refusal score
 AI-shoring alignment: early attempts to scale AI safety demonstrate promise
AirStreetCapital · case_study
AI-shoring alignment: early attempts to scale AI safety demonstrate promise
42% — win rate
 Models are capable of faking alignment...😬
AirStreetCapital · case_study
Models are capable of faking alignment...😬
 ...but many models do not fake alignment at all
AirStreetCapital · diagnosis
...but many models do not fake alignment at all
5 — Alignment faking prevalence
 Despite breakthroughs, there is still no fool-proof mitigation for scheming
AirStreetCapital · diagnosis
Despite breakthroughs, there is still no fool-proof mitigation for scheming
30x — Covert behavior rate
 Unexpected generalization: narrow fine-tuning can unlock a “cartoon villain” persona
AirStreetCapital · case_study
Unexpected generalization: narrow fine-tuning can unlock a “cartoon villain” persona
 ...but this could actually bode well for alignment science
AirStreetCapital · case_study
...but this could actually bode well for alignment science
 LLMs can read between the lines
AirStreetCapital · diagnosis
LLMs can read between the lines
 Could training data create self-fulfilling misalignment?
AirStreetCapital · diagnosis
Could training data create self-fulfilling misalignment?
 Subliminal learning: LLMs pass on traits via hidden signals in data
AirStreetCapital · other
Subliminal learning: LLMs pass on traits via hidden signals in data
 Early applications of attribution graphs reveal internal mechanisms
AirStreetCapital · context
Early applications of attribution graphs reveal internal mechanisms
 Personality engineering with persona vectors
AirStreetCapital · diagnosis
Personality engineering with persona vectors
 From filters to fortresses: prompt injection defense gets architectural
AirStreetCapital · case_study
From filters to fortresses: prompt injection defense gets architectural
100% — prompt injection success rate
 Gradual disempowerment and the “intelligence curse”
AirStreetCapital · context
Gradual disempowerment and the “intelligence curse”
 Mitigations for open-weight models: useful friction, not a solution
AirStreetCapital · key_messages
Mitigations for open-weight models: useful friction, not a solution
 Paths forward: 1) Deterrence and non-proliferation, just lock it down
AirStreetCapital · context
Paths forward: 1) Deterrence and non-proliferation, just lock it down
 Paths forward: 2) Adaptation buffers, building resilience over restriction
AirStreetCapital · strategic_options
Paths forward: 2) Adaptation buffers, building resilience over restriction
 Paths forward: 3) Implement science-first policy
AirStreetCapital · framework_other
Paths forward: 3) Implement science-first policy
 OpenAI and Anthropic test each other’s models on safety evals for the first time
AirStreetCapital · case_study
OpenAI and Anthropic test each other’s models on safety evals for the first time
 China turns up the heat on AI Safety
AirStreetCapital · industry_trends
China turns up the heat on AI Safety
3,500 — Number of papers
 ...yet China's safety practices have not fully converged with the West
AirStreetCapital · industry_trends
...yet China's safety practices have not fully converged with the West
 Our survey of 1,183 participants reveals significant AI usage and productivity gains
AirStreetCapital · situation_overview
Our survey of 1,183 participants reveals significant AI usage and productivity gains
1,183 — survey respondents
 >95% use AI at work and in their personal lives, and 76% pay out of their own pockets
AirStreetCapital · key_takeaways
>95% use AI at work and in their personal lives, and 76% pay out of their own pockets
76% — Monthly AI spend
 92% of respondents report increased productivity gains from gen AI services
AirStreetCapital · key_takeaways
92% of respondents report increased productivity gains from gen AI services
92% — Productivity gain
 Users look to AI for productivity, coding, and research...often replacing traditional search
AirStreetCapital · key_takeaways
Users look to AI for productivity, coding, and research...often replacing traditional search
45% — Share of respondents
 What was the most surprising moment you had in the last year with AI?
AirStreetCapital · key_takeaways
What was the most surprising moment you had in the last year with AI?
28% — Share of respondents
 Hot or not? Which AI tools have you started vs. and stopped using this year?
AirStreetCapital · industry_trends
Hot or not? Which AI tools have you started vs. and stopped using this year?
59 — Number of times mentioned
 AI services are largely run directly from OpenAI and Anthropic or hyperscalers
AirStreetCapital · industry_trends
AI services are largely run directly from OpenAI and Anthropic or hyperscalers
44.1% — Respondent preference
 Users sort of care for the location of their AI datacenters, but won’t switch because of it
AirStreetCapital · key_takeaways
Users sort of care for the location of their AI datacenters, but won’t switch because of it
86.3% — Respondent percentage
 >70% report their organization’s budget for gen AI to have grown in the last year
AirStreetCapital · kpi_dashboard
>70% report their organization’s budget for gen AI to have grown in the last year
70% — Budget growth
 The AI regulatory landscape has not significantly impacted AI strategies...so far
AirStreetCapital · industry_trends
The AI regulatory landscape has not significantly impacted AI strategies...so far
27.8% — Survey response percentage
 The most frequently used gen AI use cases within organizations
AirStreetCapital · market_landscape
The most frequently used gen AI use cases within organizations
72.2% — Percentage of respondents
 ChatGPT, Claude, Gemini/Google and Perplexity are used most regularly
AirStreetCapital · kpi_dashboard
ChatGPT, Claude, Gemini/Google and Perplexity are used most regularly
82.6% — Respondent usage frequency
 Developers love Cursor, Claude Code and GitHub Copilot
AirStreetCapital · competitive_analysis
Developers love Cursor, Claude Code and GitHub Copilot
46% — Developer preference percentage
 Outside of developer tools, which AI services are most popular?
AirStreetCapital · market_landscape
Outside of developer tools, which AI services are most popular?
129 — Number of times mentioned
 AI is mainly procured through APIs, followed by fine-tuning and building from scratch
AirStreetCapital · key_takeaways
AI is mainly procured through APIs, followed by fine-tuning and building from scratch
71.8% — Respondent preference
 Whether AI runs on public/private/on-prem, at the end of the day, it still uses a GPU
AirStreetCapital · analyze_data
Whether AI runs on public/private/on-prem, at the end of the day, it still uses a GPU
84.7% — Respondent usage percentage
 How 1.2k practitioners rated their top AI labs
AirStreetCapital · peer_benchmark
How 1.2k practitioners rated their top AI labs
1.2k — Practitioner rating
 Join our global community of best practices events (airstreet.com/events)
AirStreetCapital · next_steps
Join our global community of best practices events (airstreet.com/events)
 €14bn of customer value at risk due to customer experience
misc · 2018 · problem_statement
€14bn of customer value at risk due to customer experience
€14bn — net margin
 Poor customer experience has strong impact on client value
misc · 2018 · problem_statement
Poor customer experience has strong impact on client value
€14BN — Customer margin base at risk
 CVaR is calculated as the probability of decreasing relation because of bad experiences times the expected loss on value
misc · 2018 · diagnosis
CVaR is calculated as the probability of decreasing relation because of bad experiences times the expected loss on value
 In the four largest European countries, we estimate the total CVaR to be €14bn
misc · 2018 · market_sizing
In the four largest European countries, we estimate the total CVaR to be €14bn
€14bn — CVaR
 Digital leaders are best performers and biggest entities are typically at or below average
misc · 2018 · benchmark_peers
Digital leaders are best performers and biggest entities are typically at or below average
0-10 — CX satisfaction score
 Despite the metric used to monitor CEX, Digital banks perform better
misc · 2018 · benchmark_peers
Despite the metric used to monitor CEX, Digital banks perform better
Next page →