←
Chris Ackerson102 posts
Chris Ackerson avatar

Chris Ackerson

@cackerso

SVP Product AlphaSense

◷ Joined May 2009▧ 7 Photos and videos
Chris Ackerson avatar
Chris Ackerson@cackerso·Nov 13

How does an always-on agent work 24/7 without its bill growing exponentially? It's the system around the model that enables SuperAnalyst to stay on the frontier while working continuously on your behalf: - High quality content (signal > noise) - Search is core, not an afterthought - Pre-computed context (most exciting new lever- give the agent 'experience' and 'intuition') - Updates pushed in real time, so the agent acts only on change - A frontier research sub-agent that brings context to the agent, instead of the other way around. I'm most excited about SuperAnalyst as infrastructure - fleets of SAs, managed programmatically, monitoring markets, and executing orders of magnitude more and different work than was ever possible before. We're all going to need more GPUs.. This is becoming the reality inside AlphaSense (who do you think pre-computes the context?) and soon available to everyone via our SuperAnalyst API. Details here: alpha-sense.com/resources/te…

Chris Ackerson avatar
Chris Ackerson@cackerso·Nov 4

Interesting framing - this is exactly how we designed our SuperAnalyst MCP. Think of it as an intelligence engine efficiently producing exactly the right context for any investment research use case, continually optimizing itself in real time. Context is the true bottleneck to greater AI adoption in finance. Let your agent point its tokens on the problem you’re solving, not on extracting data from hundreds of sources and millions of documents and data points. alpha-sense.com/resources/pr…

I much prefer the concept of bringing context to the agent vs. bringing the agent to the context. What do I mean by that? Rather than asking an agent to discover its own context through tool calls, file exploration, and RAG queries… Build a pipeline that understands your agent use cases and arranges for the relevant context to be materialized just in time and kept up to date. Then have your harness match each use case to the context it needs.

Chris Ackerson avatar
Chris Ackerson@cackerso·Nov 4

Today @AlphaSenseInc, we're launching the SuperAnalyst MCP,  putting SuperAnalyst to work wherever your agents run. Rather than one more data pipe for your agents to brute force there way across, SuperAnalyst MCP allows them to delegate research to the world's best research system - thousands of sources, market-leading search, purpose built AI - so your agent spends its tokens on the problem you actually want solved. We measure dramatically higher quality at 3x lower cost with ever claim cited to its source down to the sentence. SuperAnalyst has become the platform for technology teams to automate complex investment research and build institutional grade market intelligence applications. Highly encourage folks to check it out: alpha-sense.com/resources/pr…

Andrew Feldman avatar
⟳ Chris Ackerson retweeted
Andrew Feldman@andrewdfeldman·Oct 30

Interesting look under the hood of @AlphaSenseInc's SuperAnalyst. Lots of thoughtful engineering in how they give agents the tools to do useful research while keeping access tightly controlled. And when speed matters, they route requests to open-weight models on Cerebras. Worth a read if you’re building agents:

When an AI agent escapes the sandbox, the first question should be: who left the door open? For firms in regulated industries, giving agents broad access to data and tools has meant weighing what they can do against what they might reach. SuperAnalyst runs every session in its own isolated MicroVM, with scoped credentials, a sealed network, and identity checks at each step, so one failed control never becomes a breach. Lawrence Sica, VP of Product Security, and @chaschwartzz, Staff Software Engineer, walk through each layer of SuperAnalyst’s sandboxing architecture and how they work together to keep agent workloads contained. Read the full article: alpha-sense.com/resources/te…

Post attachment
Chris Ackerson avatar
Chris Ackerson@cackerso·Oct 26

SuperAnalyst driving another acceleration in token consumption at AlphaSense given step change in complexity of work automated and executed. Product work now is in teaching all users how to harness that power.

Updated Token Consumption Tracker For Anthropic S-1 day, breaking out the surge in Ant tokens based on usage on AlphaSense Update your models accordingly

Post attachment
Chris Ackerson avatar
Chris Ackerson@cackerso·Sep 13

The obvious way to bring both sides together is to allow distillation for open weights models by American companies.

We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three-part plan for doing so. Anthropic is unilaterally committing to the first of these steps. We’ll provide third-party evaluators with permanent, employee-level access to our systems, so that they can verify adherence to our safety measures, report on incidents, and assess models’ alignment during training. You can read the full post here: darioamodei.com/post/we-must…

Chris Ackerson avatar
Chris Ackerson@cackerso·Aug 10

Meta being back in the open weights game is a very big deal. There have been 3 major training (SFT/RL) periods since the ChatGPT moment: 1. 2023 when the closed models weren't good enough for many use-cases and weren't enterprise ready (ZDR, security, latency, hyperscaler support). AlphaSense was big on fine-tuning (including on Llama 2/3) in this period and it helped us launch features like Smart Summaries at sufficient reliability. But as closed models accelerated (Sonnet 3.7, GPT-4o and especially reasoning models like o1) the market including AlphaSense re-focused around the labs to power features like Generative Search and Deep Research (still maintained custom models for narrower data processing and ranking use-cases). 2. 2025 when the labs released their fine-tuning endpoints and a number of startups hyped these offerings. This was always a flash-in-the-pan and we didn't take it seriously. 3. Right now. The ecosystem around training is fully here with data generation, RL infra and American open weights models all feeding off each other. Once again at AlphaSense, we see a massive opportunity to increase the quality, reliability and efficiency of our AI offerings by training custom models. In particular we are very focused on models for Search - specifically context collection across our qualitative and quantitative tools and data which continues to be the biggest bottleneck and opportunity for progress in AI in Finance.

Today we're also opening the weights for Muse Glimmer, a great 30B parameter dense model that can run locally. Soon we'll also release the weights for Muse Spark 1.2, our latest foundation model. Meta is a strong supporter of open source and I'm proud of these releases. Congrats to @alexandr_wang and the MSL team for all your great work on these models.

Chris Ackerson avatar
Chris Ackerson@cackerso·Aug 6

This is correct and we are already seeing glimmers of recursive self-improvement from internal use accelerating SuperAnalyst development. But @AlphaSenseInc we are taking this idea even further. In Finance, a growing share of our workload leverages recurring structures: supply-chain relationships, channel check signals, and the key debates around companies, industries, and cross-sector themes. Today, an agent reconstructs these from hundreds of documents, per user, per query, at run-time. Once you see agent output as a knowledge asset, a strategic question follows immediately: why wait for the request? We're deploying agents that pre-compute them as first-class, queryable datasets: supply-chain graphs assembled across filings, transcripts, and expert calls; channel-check syntheses coordinated by AI interviewers; key-debate ledgers maintained per ticker with the bull and bear evidence attached, updated as new documents land. Our SuperAnalyst starts from the pre-computed structure and spends inference only on the delta, each new question, new document, new angle leading to better consistency, lower latency, and cleaner citations as side effects. Just like standardized fundamental data helped analysts avoid extracting the same data manually from unstructured company disclosures, we are now structuring the next major layer of analysis that everyone has been doing independently.

Knowledge flywheels are a new scaling dimension for self-improving AI.

Chris Ackerson avatar
Chris Ackerson@cackerso·Aug 5

And the biggest lever: the most efficient token is the one never spent. We're pre-computing supply chains, channel checks, and key debates as first-class datasets so SuperAnalyst starts from structure and spends inference only on the delta. Orders of magnitude impact.

Chris Ackerson avatar
Chris Ackerson@cackerso·Aug 5

Next: we're RL-training our own search agent on this eval machinery: rubrics + judges + millions of trajectories. Target: model routing-level quality at up to 40x lower cost. Frontier models aren’t going anywhere: they keep the reasoning, coding, analysis. We specialize in the context.

Chris Ackerson avatar
Chris Ackerson@cackerso·Aug 5

One deliberate constraint: every config ran on the same indexed corpus. None of this measures the content itself: proprietary expert call transcripts, broker research, events, journals. These gaps are a floor.

Chris Ackerson avatar
Chris Ackerson@cackerso·Aug 5

The methodology: because benchmarks are easy to game, 245 complex multi-step questions with expert rubrics written before any model answered. Hard requirements on what matters to professional users. We update and run these benchmarks continuously: every major model release, every major feature update, month after month, against increasingly complex, real-world tasks.

Chris Ackerson avatar
Chris Ackerson@cackerso·Aug 5

No single model wins most of the time and different models excel at different question types: e.g. metric extraction, screening, long-form synthesis etc. Route each task to the right model and you reach 2.8x baseline.

Post attachment
Chris Ackerson avatar
Chris Ackerson@cackerso·Aug 5

Open weights will play a critical role: Gemma 4-31B matches Claude Sonnet 5 on answer quality at ~40x lower cost and inference on @CerebrasSystems runs 10-20x faster. Good-enough quality at low cost and latency unlocks lots of exciting use-cases.