Mullet loving woodchuck skinning model merge of AI, search, and redneck. I refuse to believe vector retrieval is the final form of how systems get context.
NY
Joined March 2018
- Tweets1.7K
- Following394
- Followers814
- Likes1.3K
Never has there been a product that has as much PMF but suffered such a quick demise as deep research. Everyone used it. Everyone loved it and now the term is dead.
The more I work on an agentic product, the clearer it becomes that things are complicated and will only get more complicated. I added some keep-cache-warm calls to help with subagents and when users come back and ask questions after cache expiry. Saves ~3% and keeps it snappy.
Daniel Campos retweeted
My Investment Memo builder in SuperAnalyst is now 🔥 – beats out what I was doing as a HF analyst. I’m looking for people that want to collaborate to draw smaller and smaller circles.
It uses a Canalyst model on my behalf. It reasons to assumptions (depending on type of company, builds out TAM and penetration, models cyclicality, etc), researches what experts are saying about key debates, builds to a valuation framework taking into account Comps / multiple history, etc. Full companion model / backup in excel.
Roll this forward a bit and it’s obvious HFs will be built trading on the back of something like this (much higher volume of getting to decisions faster, neutrally). I’d especially love to collaborate with some of you building AI HFs.
DM if interested – must have AlphaSense account (the skill calls our tools)
As we scale always-on agents at @AlphaSenseInc SuperAnalyst, a natural question arises: what model do users choose for their interaction and orchestration? Most users trust the platform default (~70%), with the remaining mostly split across Opus 5.5 and Astra.
Some of the most useful workflows in finance are those that track and quantify key metrics and views over time. The downside is that these can be incredibly tokenhungry and slow to make. We took our wide corpora and turned it into composable blocks you can assemble and understand
We’ve officially launched Channel Check Intelligence
The best new source of Idea Gen
AlphaSense began structured channel checks a bit over a year ago, already scaled to 3k per month (compliance reviewed, no embargo). Already much more than anyone could read but high signal if you can make sense of it.
Channel Check Intelligence makes sense of it —> what sectors, companies are inflecting, pricing / invy better or worse etc. Not only the numbers, but the why.
CCI is another form of pre-compute where we use continually use LLMs to structure the data once, for everyone. Soon, when you build your Earnings Previews with SuperAnalyst, the system will know to bring in CCI - and even use as an input to the assumptions in your model - without needing to rebuild the analysis from scratch (accuracy, time, tokens)
We are investing aggressively in scaling and development -- it's going to get pretty crazy in a year.
I get genuine joy from products that read my mind before I explicitly ask for it. I finally got it working perfectly in the @AlphaSenseInc SuperAnalyst. A simple request for a briefing gives me the most information-dense point of view and my requested East London rap tone.
Working with always-on harnesses like Muse, Instinct, Hermes, SuperAnalyst(@AlphaSenseInc plug) really highlights the role of memory and context and their relation to inference. I expect the model to know everything, but I don't want to have massive cold starts with long context
I did a deep dive into my usage. Assume a 5-minute cache TTL. Across all these surfaces, my usage is out of cache >70% of the time. Moreover, when I come back 50 minutes later, it's not clear that the agent needed to know how it did tool calls at the beginning of the session.
Anthropic finally stopped giving its internal models full access to the internet. Before that the response was essentially "OMG we gave the model a gun and it shot people." Just don't give it a gun.
I genuinely don’t know if I say this enough, but I really fucking love America best country in the world if you disagree, you can fuck off
I dont know about you but I like getting more for less. Save time, get better F1. Datology. AI data curration made easy (Read it in the TBPN voice)
Haiku 5.5 is out, and if you know me, 4.5 was one of my best friends. At @AlphaSenseInc we have been testing it for the last few weeks and are happy to report it's a great model and a no-brainer replacement for all your scale- and cost-constrained use cases!
I wanted to understand how AI systems are changing user behavior @AlphaSenseInc so I dove into the last 2 years of query behavior. TLDR users ask for deliverables instead of documents now
What's next isn't "delegate everything." In SuperAnalyst's alpha month, delegation barely moved, from 56% to 58%. But prompts are 50% longer, and 39% need 3+ capabilities, compared to 29% before. People aren't handing off more. They're asking for harder things.
The first two years of this AI wave moved work from the person to the machine. The next two years are about what people build when the machine can take on much harder work. Full write-up: alpha-sense.com/resources/te…
