47 talks · Two days · Source-linked analysis
Find the sessions behind a topic. Compare how widely it appears, see what changes between days, and check the source.
These are overlapping term families, not rankings of importance. The field guides explain what to try.
Cost · 34 of 47 talks across both days
Scroll the table to compare topics. On a phone, topic coverage offers a simpler way to find the same sources.
Talks with matches in at least two transcript segments.
Both days · 34 of 47 talks
Matches in at least two transcript segments per talk. This measures coverage, not importance or agreement.
Read the related field guide →Labels describe these passages. Results remain speaker-reported.
In terms of cost, in terms of impact, self-hosting GLM, our blended cost of token is 30 cents per million, and so our invoice is one-tenth of what we were spending on.
The next segment names Anthropic and OpenAI as the comparison. This is a speaker-reported invoice result, not a total-cost benchmark. Hardware, staffing, and workload equivalence are not established here.
Day 1 · David Abitbol
Session notes · Watch this passage ↗
We need to add the AI spend and the human spend, the human salary spend.
The next segment completes the point: calculate the real cost per pull request. This is a measurement recommendation.
Day 2 · Steve Jamison
Session notes · Watch this passage ↗
Today, price per token is a useful practical proxy, especially for comparable workloads and quality requirements, and gets us closer to delivered output, much closer than the often quoted, all you finance bros out here will have heard this, GPU dollar per hour price, or as we view it, a cost.
The speaker calls token price a practical proxy. The preceding passage says price per successful task is the goal and is difficult to standardize.
Day 2 · Oliver Kennon
Session notes · Watch this passage ↗
These automatic matches show why a talk is counted. Read the context before interpreting a match.
Day 1
This was built on the foundation that advisors only spend approximately 20% of their time with clients.
Day 2
Over the last year, the number of use cases that use LLMs has increased 123%, while at the same time, a lot of our top features are growing in token spend, but on a unit economic basis, they're actually becoming cheaper because there are frontier models that are being released all the time now, and these models are cheaper, faster, and better.
Categories overlap. Broad terms can have different meanings in different talks. A match does not establish the talk’s main subject.
Exact matching rule (case-insensitive regular expression):
\b(?:cost\w*|prices?|pricing|spend\w*|budgets?|expensive|cheaper?|dollars?)\bBoth days stay visible here. The day filter controls the coverage list and source sessions above.
Matches per 10,000 words · Pool all words in each day. Longer talks contribute more words.
Breadth: Day 1 matches 14 of 23 talks; Day 2 matches 20 of 24. Rates count all matches, including talks below the coverage threshold. This is a descriptive comparison of two captured streams.
Word counts explain the vocabulary. They do not measure usefulness. Inflections and common aliases are merged; common function words are removed.
These counts cover both days and do not follow the day filter. “Agent” includes “agents” and “agentic.” Lexical term counts and topic-family counts use different matching rules.
The corpus contains 134,250 words from 47 distinct talks in the two main-stage recordings. It does not cover every room.
Semantic topic models and sentiment scoring were tested during research. They are not presented as definitive topic labels or speaker attitudes. Technical threat descriptions and ordinary words such as “save” can mislead a general sentiment model.
Coverage data (CSV) · Word counts (CSV) · Explorer data (JSON) · Method and source hashes (JSON)