417 results · Most relevant
Anekha describes how Moshi combines public and private data, cleans and transforms it, applies statistics and machine learning, builds an ontology and n-dimensional graph, and delivers software that pushes relevant data to users. She frames the volume of processing as beyond human capacity, citing that in the time it takes to say a sentence Moshi has learned 300 new things.
Anekha breaks the market into structured data (futures time series, technical indicators, moving averages, fundamentals time series with demand and supply data) and unstructured data (news, X/Twitter data, podcast audio). She also describes a third category: the user's private data, such as analysts they follow, subscriptions, and internal personalization data like LNG ship counts through the Strait of Hormuz.
The speakers describe the Twitter/X firehose as an extremely expensive alternative data source — roughly $300,000 per month for a hedge fund to ingest all tweets — which early buy-side firms like Balyasny were among the first to use; it was free for researchers. They caution that social media data is very noisy because people talk nonsense and some deliberately try to manipulate markets, and Anekha Sokhal says she would not trade on it without filtering.
The $300,000/month figure is approximate ('something like') and carries a possible_numeric_error provenance flag; the free-for-researchers point is stated with lower confidence.
You can try to change how a calculation works and you're likely going to get some pushback from the harness. If you change multiplication for an aggregation, make the quick edit, and save the file, you sabotage the analysis — you won't be able to reach any conclusion of the analysis you wanted, because the harness is going to stop you by its own definitions and the way it was built.
Experts often do not check every cell; working with finance teams in data, they could open a spreadsheet and look at a few cells they know will give a ballpark total. If the ballpark total doesn't feel right, experts will let you know — e.g., 'how did we spend this much money on marketing, I don't think that's right' — and it's a little bit of a gut reaction.
The speaker spent a month and a half building a harness to avoid resizing text boxes in Google Slides. The harness dug into problems and gathered documents, but when he gave an example for a modification, the model fixated on it and put it everywhere in the presentation. He ended up showing a colorful example to senior leadership who asked where the number and assumption came from — and he didn't know.
A speaker with a semi-strict monthly budget close left categorization, aggregates, refunds, and expenses to the agent, found they were getting slightly different results every time, and spent more money and time wrangling with the agent to get results it had gotten right three times but not the fourth — so they decided the budget calculations needed to be a script instead of agent-generated code.
A person used a shared 'master file' that contained hidden assumptions about reasonable value ranges. The file changed without their knowledge; they did nothing wrong, but their results were entirely wrong. This illustrates how lack of traceability creates blind spots when processes change.
Focus on aligning on what outcome you're trying to drive with AI rather than gaining trust on the AI implementation itself.
endorsed advice
Whatever you design, if you have a feedback mechanism, make sure you tailor it to how you actually want to use it.
Anekha states you want to make sure you have agreement on how a third party is processing your data and what they're doing with it.
The hosts present the scenario of a data person or analyst saying 'no, no, no, here's the result. Trust me, I ran the numbers' — and argue you should not trust that.
The hosts warn that if you don't have a way of having traceability even for changes you make in your code, you're leaving yourself vulnerable to blind spots on the process change.
Golden examples chosen for a domain are based on the picker's biases and selection bias; applied to a slightly tangential domain they may turn out to be nonsensical. Golden examples give ground truth for staying in the right space, but no domain is isolated.
endorsed interpretation
You keep accountability high by being able to see under the hood, inside the black box. Accountability mechanisms in harnesses also reduce the stress of wondering in your head whether the agent made a wrong assumption.
Not all information is the same — something that bites data professionals on a weekly basis. If you present an analysis or do financial analysis on a potential investment, you need to show where your numbers came from and attribute your decision to the body of evidence you had. Treating everything as a fact is a mistake to avoid — some inputs are estimates, not hard truths.
The whole framework and all the principles started from trying to get to that shared domain, shared context — you're essentially translating your language into the training dataset for the model. You work with the model to say: this is how I understand the problem or how my company does this — and the model responds: this is how I fit it in my training dataset.
Felippe Felisola Caso recommends applying AI deliberately and tying it to concrete business value rather than token maxing, and applying it to the specific things you care about in your life rather than setting up a general instance connected to everything.
Felippe warns that if you put a gate before every decision asking 'ask me', it becomes very low signal again.
Francesco Galletta recommends getting a second person to review your ideas before implementing them, admitting he is guilty of thinking he has the best idea in the world only to have Felippe tell him not to implement it.
No published protocol matches this selection.
The creators behind most of what cleared.
344 cleared insights match
71 cleared insights match
Choose sources → to get a library of your own and a daily brief.
