Third-year PhD candidate in the MIT Data Systems Group. Working on Palimpzest and semantic operators. Project: https://nitter.cf/t.co/2gSZUQjfaQ Email: mdrusso@mit.edu
Boston, MA
Joined May 2024
- Tweets28
- Following78
- Followers71
- Likes69
Here here
Lots of disdain in the comments. It’s honestly a bit unnerving to read people’s negative reaction to a harmless plea to continue prioritizing sharing discoveries and results, as humanity has done for millennia. Maybe it’s easy to think that the authors are operating from a place of ego and jealousy, if they are Fields medalists, and therefore the general public doesn’t trust them. So let me try explaining as someone who has no career achievements, no stake in the frontier labs, and no jealousy because my peers are all early-career like me 😆
Academics are stewards of history. We have 3 jobs. (1) to train people / teach them what has been done in the past. (2) we make some discoveries ourselves, and (3) we regularly interact with our peers to learn about their discoveries; what tomorrow’s history may be. Many people think we just do (2), but in fact all 3 are important; otherwise we would be highly inefficient in problem-solving over the long term (centuries, millennia).
We cannot really make AI the stewards of history (there are capitalist and geopolitical incentives not to). If we rid the world of academic values, there may not be immediate consequences, but things will be very bad in the long term. Progress will stall. Eventually folks will realize we need a clear and continually-updating ledger of history (ie “memory” for the AI-pilled folks). It will take a lot of work to rebuild this culture from first principles
On the optimistic side I think academics have a golden opportunity to figure out how we can harness AI in pursuit of our values. We can probably keep more information in our history, disseminate info more broadly, review and validate each others’ discoveries faster, etc. But, given that academics are powerless compared to these trillion dollar companies, we can only do the best we can—speak up; discuss; educate the public about academic values (which we haven’t really had to do since ww2); and, honestly, pray
Letting agents naively query your enterprise data lake can be error-prone and expensive. What if we could structure their execution in a way that was steerable and easy to optimize for cost-efficiency? This week I’ll be presenting two research papers at VLDB which work towards this goal.
#Carnot: vldb.org/pvldb/vol19/p4642-r…
#Abacus: vldb.org/pvldb/vol19/p1060-r…
#Carnot is a system for interpretable, interactive, and optimized execution of agentic search queries. Carnot lets business users issue queries in natural language, compiles them to an optimized computation graph, and executes them. A notebook interface gives users control over the agent’s query, enabling them to inspect intermediate outputs while steering the agent back to the correct computation when it makes a mistake. Finally, Carnot helps users manage the cost of their queries by predicting the cost of each operation and opting for plans which can satisfy a user’s specified budget.
#Abacus is an optimizer for semantic query processing which we built and shipped in Palimpzest. Abacus optimizes semantic queries (with or without relational operations) across all three dimensions of quality, cost, and latency. The optimizer was designed to be extensible and scalable, leveraging a multi-armed bandit search algorithm to find the best operators for each query plan and optimization objective. In our evaluation, Abacus' plans achieved ~7%-40% better quality while being 10.8x cheaper and 3.4x faster (on average) than the next best system(s).
This work is the product of a great research team from MIT’s Data Systems Group and OASYS Lab: Yash Agarwal, @tianyu_li_, @astrogu_, @chunwei_l, Siva Sudhir, @gerarvita, @MikeCafarella, @lateinteraction, @tim_kraska, and @samrmadden
If you’re at VLDB this week and curious about our current (and future) work, please swing by our demo, research talk, and poster sessions!
Who doesn’t love a friendly competition between AI systems researchers?
Very excited to be repping MIT alongside @tianyu_li_ @astrogu_ @gerarvita at the first ever Grounded Reasoning Cup!
Introducing the lab sponsors for the Databricks Grounded Reasoning Cup at #DataAISummit 2026: @AnthropicAI, @OpenAI, and @GoogleDeepMind.
Each lab is partnering with leading academic teams to build agents that tackle grounded reasoning over complex government data using the latest models and tooling.
Meet the teams, see the agents they’re bringing to the competition, and join us as they push the boundaries of enterprise AI reasoning live on stage! 🏆
Recent agentic systems (Claude Code, Codex, RLM, etc.) push context out of the prompt and into the environment (e.g., as files). This helps them maintain long-term knowledge about their goals and functionality.
🚨 While this is a good idea, we show a surprising result: systems that use external environments like this perform much better when given a small, fixed-size, in-context, agent-managed cache that "𝘱𝘦𝘦𝘬𝘴 𝘪𝘯𝘵𝘰" these environments.
🚀 Our paper, 𝗣𝗘𝗘𝗞: 𝙖 𝙨𝙮𝙨𝙩𝙚𝙢 𝙛𝙤𝙧 𝙗𝙪𝙞𝙡𝙙𝙞𝙣𝙜 𝙖𝙣𝙙 𝙢𝙖𝙞𝙣𝙩𝙖𝙞𝙣𝙞𝙣𝙜 𝗮𝗻 𝗼𝗿𝗶𝗲𝗻𝘁𝗮𝘁𝗶𝗼𝗻 𝗰𝗮𝗰𝗵𝗲 𝙛𝙤𝙧 𝙇𝙇𝙈 𝙖𝙜𝙚𝙣𝙩𝙨, introduces this idea.
Compared with strong baselines, including RAG, Compaction Agents, and SOTA prompt-learning frameworks, PEEK dominates the cost–quality Pareto frontier: achieving +6.3–34.0% in quality, with fewer iterations and lower cost.
Paper: arxiv.org/abs/2605.19932
GitHub: github.com/zhuohangu/peek
More in the thread below! (1/N)
Matthew Russo retweeted
Congratulations to the 2026 @cidrdb prize awardees!
@tianyu_li_ → Gong Show Winner
@FuhengZ → Database Quiz Winner
They each received a rare signed print of "The Birth of the Database Messiah" (est value $12,000).
ALT Tianyu Li (Wisconsin) CIDR 2026 Gong Show Winner
ALT Fuheng Zhao CIDR 2026 Database Quiz Winner
Matthew Russo retweeted
I am SUPER EXCITED to publish the 131st episode of the Weaviate Podcast with Matthew Russo (@RussoMatthew), a Ph.D. student at MIT! 🎉
AI is transforming Database Systems. Perhaps the biggest impact so far has been natural language to query language translations, or Text-to-SQL. However, another massive innovation is brewing. 💥
AI presents new Semantic Operators for our query languages. For example, we are all familiar with the WHERE filter. Now we have AI_WHERE, in which an LLM or another AI model computes the filter value without needing it to be already available in the database!
```sql
SELECT * FROM podcasts AI_WHERE “Text-to-SQL” in topics
```
Semantic Filters are just the tip of iceberg, the roster of Semantic Operators further includes Semantic Joins, Map, Rank, Classify, Groupby, and Aggregation! 🛠️
And it doesn’t stop there!
One of the core ideas in Relational Algebra and its influence Database Systems is query planning and finding the optimal order to apply filters.
For example, let’s say you have two filters, the car is red and the car is a BMW. Now let’s say the dataset only contains 100 BMWs, but 50,000 red cars!! Applying the BMW filter first will limit the size of the set for the next filter! 🧠
This foundational idea has all sorts of extensions now that LLMs are involved! This opportunity is giving rise to new query engines with declarative optimizers such as Palimpzest, LOTUS, and others! 💻
So many interesting nuggets in this podcast, loved discussing these things with Matthew, and I hope you find it interesting! 👇
📣 We're spreading the word about #SemBench -- a brand new benchmark for semantic query processing over multimodal workloads including text, image, audio, and tabular data!
📜Paper: bit.ly/3WGsZf6
💻Website: sembench.org
💾Code: bit.ly/49DeuAg
👀 See Immanuel Trummer's original announcement of #SemBench: linkedin.com/posts/immanuel-…
If this (very high-level) summary of our work has piqued your interest -- go read our full paper!
📄Paper: arxiv.org/pdf/2405.14696
💻Code: github.com/mitdbg/palimpzest…
We would love to hear any feedback, ideas for more use cases, and/or opportunities for collaboration.
Finally, this work is the product of a great research team:
Chunwei Liu (@Tranway), Michael Cafarella (@MikeCafarella), Lei Cao, Peter Chen, Zui Chen (@ZuiChen), Michael Franklin (@franklinmj), Tim Kraska (@tim_kraska), Sam Madden (@samrmadden), Gerardo Vitagliano (@gerarvita)
For our evaluation, we curated three SAPP workloads in:
- Legal Discovery
- (identifying evidence of fraud at Enron)
- Real Estate Search
- (finding a suitable home in Cambridge, MA)
- Medical Schema Matching
- (reproducing a real-world bioinformatics paper)
Second, using three different policies on each dataset, we find that PZ selects high-quality plans.
Plans selected by PZ consistently have similar or better quality than naive GPT-4 baselines, with up to 80.0% lower single-threaded runtime and up to 89.7% lower cost.
First, we find PZ produces appealing plans at numerous points in the tradeoff space.
PZ produces plans that are:
- 4.7x faster, 9.1x cheaper, and within 85.7% of the naive GPT-4 plan's F1-score
- 3.3x faster, 2.9x cheaper, and up to 1.1x better F1 than the naive GPT-4 plan
