12 live patterns across four podcasts
Receipts for what people say.
Savlar listens across long-form conversations and shows who agrees, who disagrees, who changed their mind and who predicted what, each claim linked to the second it was said. So you and your AI can finally cite it.
“That one of the big challenges is having doing effectively a sensing part in a reasonable way which also gives you relatively strong guarantees because if you, first of all, you don't get strong guarantees, you might crush someone.”
“It would be a lot more unstable. You know, safety will be extremely, extremely hard to guarantee.”
Expected difficulties in guaranteeing robot safety
3 speakers · 2 shows · 3 claims
The tracked speakers share an expectation that robot safety assurances will be difficult: Alexander Mattick warns that a robotic system without strong guarantees might crush someone. Open the pattern
Caveat. These are related concerns, not identical predictions: the speakers do not all predict crushing injuries.
- Inferred
- Jason Ma’s claim supports Alexander Mattick’s · confidence 0.85, above the 0.85 it needs
- Concluded
- A shared prediction · confidence 0.85
For anyone who has to quote what was said out loud: analysts, reporters, researchers, and the agents they use.
Or ask your own agent:
claude mcp add --transport http savlar https://That line is for Claude Code. Using Claude, ChatGPT or Cursor? Every step
Listening to
- Dwarkesh Podcast
- Latent Space: AI Engineering Podcast
- Machine Learning Street Talk (MLST)
- No Priors
- episodes heard
- 33
- hours of audio
- 46
- claims, each with its moment
- 11,183
01 The problem
You can’t search it, compare it, or quote it with proof.
Most of what matters is said out loud. Nobody can cite it.
- Earnings calls
- Debates
- Hearings
- Town halls
- Meetings
- Interviews
- Podcasts
- Lectures
Summaries lose who said what.
Each recording is summarised alone, and the speaker, the exact words and the second they were said are gone.
Nobody can compare across recordings.
Who agrees with whom, who changed their story, who called it first: it lives across hundreds of hours nobody has time to hear.
AI agents put words in the wrong mouth.
Ask an assistant what someone said and it paraphrases, merges speakers, or quotes what was never said, with nothing to check it against.
So we built something that listens across all of it, and never loses the receipt.
02 How it works
From an episode’s audio to a pattern you can check, every step keeps the speaker, the words and the second.
Three steps from audio to evidence.
01 Listen.
Each episode it hears is transcribed, and each voice is named only when the system is at least 80% sure. Below that, it stays an unresolved speaker.
02 Extract.
Each claim, prediction and position is pulled out word for word, with the second it was said.
03 Connect.
Claims are linked across shows and months: who agrees, who disagrees, who changed their mind, who predicts the same thing. A pattern goes live only when it passes a fixed test, such as three speakers on two shows.
03 Watch it work
The question, the tool calls and the answer, exactly as they happened.
Four speakers. Three shows. One shared prediction.
Then your own AI finds it.
Savlar listened to all 46 hours, so nobody had to.
You
What do AI researchers expect about robots doing physical work? Give me their exact words and links.
Claude
savlar · search“robots doing physical work”done
returned 6 claims, verbatim, each with speaker, show, date and the moment
savlar · get_signal“cda785ab-992c-42dd-879a-72a15c99e626”done
returned Shared prediction · 4 speakers on 3 shows · confidence 0.85
Savlar has a live signal on this, “Predictions of AI extending into physical work through robotics”: 4 speakers on 3 shows, confidence 0.85. In their own words:
“At some point, we'll just have robots that can convert shores of silica sand and mines of copper into new computer chips.”
Dwarkesh PatelDwarkesh Podcast · 3 Aug 202610:09 “And then sometime after that, we will have robots perhaps controlled by those AIs that can do physical work in a way that broadly substitutes for humans.”
Daniel KokotajloMachine Learning Street Talk (MLST) · 8 Sep 20261:23:54 “And then, you know, like eventually the tide will wash over the physical world as well.”
Ajeya CotraDwarkesh Podcast · 1 Sep 20261:46:24
Inferred
A model judged that Kokotajlo’s prediction refines the broader one, that AI’s influence will extend into the physical world, at confidence 0.94 (bar 0.85).
Savlar’s reading: they share a broad direction, not a common forecast of timing. None of them gives a calendar date.
04 Trust
A missing pattern is better than a wrong one.
It refuses to guess.
The model named a voice “Charlie”.
That voice says: “Kudos, Charlie and Baron.”
Overruled. You can’t be Charlie if you’re thanking Charlie.
- Measured 11,183 claims, each linked to its second
- Measured 78% of speech matched to a named speaker
- The bar 80% sure, or the voice gets no name
- By rule 0 patterns without a quote behind them
A wrong name is worse than no name.
05 Not just podcasts
Each sketch shows the shape of an answer, not real data.
Podcasts came first because they’re public. The engine can run on any recording.
- Sketch
Finance
Guidance that quietly changes between earnings calls.
- Ask
- “Did the margin guidance change between the Q2 and Q3 calls?”
- Returns
- A change of mind, both moments side by side.
- Sketch
Politics
Promises, flip-flops, and who called the midterms.
- Ask
- “Who predicted the result, and when did they say it?”
- Returns
- Shared predictions, each with its date.
- Sketch
Law
Testimony that contradicts itself, with timestamps.
- Ask
- “Where does this witness contradict their earlier statement?”
- Returns
- Both statements, to the second.
- Sketch
Companies
What ten customers say in common, in their words.
- Ask
- “What do customers keep asking for on sales calls?”
- Returns
- An emerging consensus, with every quote behind it.
- Sketch
Communities
Where a town hall agrees, and where it splits.
- Ask
- “Where do residents agree on the plan, and where do they split?”
- Returns
- Consensus and disagreement, by speaker.
- Sketch
AI agents
Live todayA memory they can cite and can’t make up.
- Ask
- Any question, over MCP.
- Returns
- Exact quotes with links, kept apart from any reading of them.
Live today on podcasts, and for any agent over MCP. The other five are where the same engine goes next, with each organisation’s consent. Have recordings like these? Pilot it
06 For agents
Your agent is told to quote exactly, cite the link, and keep what was said apart from what was inferred.
Ask your own agent. It answers with receipts.
claude mcp add --transport http savlar https://- Claude Code
- Paste the line above into a terminal.
- Claude and ChatGPT
- Add a custom connector, and paste https://savlar.com/mcp
- Cursor
- Add the same address to .cursor/mcp.json
No sign-in. Six read-only tools, from get_brief to list_sources: every one on /connect
07 Work with us
Savlar is early. The fastest way in is a pilot on your own recordings.
Three ways in.
- 01Message the founder
Pilot it on your show, calls or town hall.
We set it up on your recordings and show you what it finds.
- 02Suggest a voice
Tell us whose voices to track next.
A show, a channel, a hearing: name it, and why it matters to you.
- 03Find a quote to check
Try to break it.
Report any wrong quote: every quote on a pattern’s page carries a report link.
Bilgin Koçak, founder · bilgin@savlar.com · Telegram @bilginkocak
If an answer has no link, it didn’t come from us.