BLOG

Notes on data and AI

What we are building, and what the data says when we point our own platform at a public dataset. Agentic data engineering, languages for data, model training, and case studies published with their numbers.

  1. 19
    July 23, 2026 NEW
    The Divide Inside One Country Is Wider Than Between Two

    Anthropic publishes what people actually do with Claude, 1.6 million rows a month, free. Less than half of it is work. The busiest thing it makes is an explanation, not code. And the adoption gap between two American states is wider than the gap between America and India. Built on Bonacci Studio.

    7 min readCase Studyread article →
  2. 18
    July 22, 2026
    48 Teams Were Supposed to Ruin the World Cup

    Every previous expansion was followed by a slide to the thinnest scoring in the competition's history. 2026 went to 48 teams and produced 2.96 goals a game, the highest since 1970. 104 matches, 308 goals, 1,248 players, and one pipeline that had to pull every goal out of the match it was buried in. Built on Bonacci Studio.

    6 min readCase Studyread article →
  3. 17
    June 25, 2026
    Stop Gambling on One Agent

    Run the same prompt twice, get two different answers. vibeCodeCLI races parallel fleets in your terminal instead, keeps memory across sessions, and checks the outcome before anything irreversible runs. Free, 23 providers, bring your own keys.

    7 min readProduct Launchread article →
  4. 16
    June 14, 2026
    ThinkingMemory Is Now a Memory Database for Agents

    We rebuilt ThinkingMemory around one primitive: recall. Intent in, the right context out, hybrid-retrieved, ranked, and packed to a token budget, with an in-engine lifecycle, bitemporal audit, per-tenant RLS, and an MCP server. 100% recall@5 in roughly 68% fewer tokens than dumping the corpus.

    7 min readAI Memoryread article →
  5. 15
    June 11, 2026
    Intelligence Is Deflating: 938× Cheaper in Three Years

    GPT-4-class capability cost $37.50/M tokens in 2023. Today a 2B-parameter model serves the same benchmark score at $0.04. The frontier held its price; the efficient tier collapsed. The same open model spans a 13× price spread across 12 providers. Built on Bonacci Studio.

    5 min readCase Studyread article →
  6. 14
    June 11, 2026
    Is AI Taking Jobs? I Joined the Data to Find Out

    Four public datasets (BLS/OEWS, O*NET, Anthropic Economic Index, Indeed). Three occupation coding schemes. One crosswalk. AI has not cut jobs in aggregate, but it is bending the hiring curve along the automation line before headcount moves. Built in Bonacci Studio.

    6 min readCase Studyread article →
  7. 13
    June 9, 2026
    The Current State of the AI Ecosystem: A Case Study with Bonacci Studio

    25 months of activity from GitHub, arXiv, and Hugging Face, 613 data points across five categories. RAG is up 340% year over year. Training from scratch is up 32% month over month. Nothing is cooling. The agent wrote the PySpark, ran it, and rendered the analysis live in Bonacci Studio.

    5 min readCase Studyread article →
  8. 12
    May 26, 2026
    Outcome-Aware Agents: From Token Prediction to Action Consequence Modeling

    A new research paper introducing AC-LSCM, an Action-Conditioned Latent Structural Causal Model that gives agents an explicit forward model with interventional semantics. 36× fewer safety violations than a Transformer baseline on synthetic planning tasks.

    10 min readResearch Paperread article →
  9. 11
    April 2026
    Three 9B MoE Models, and What the Compute Actually Cost

    Three 9B Mixture-of-Experts models, 122B tokens each, pretrained on Google's TPU Research Cloud. The compute bill stayed small because of the research credits. The tokenizer, the corpus pipeline, and the eval harness were the work.

    8 min readModel Trainingread article →
  10. 10
    March 12, 2026
    An Agent That Is Still Watching Your Pipelines at 3am

    It lives inside Bonacci Studio, keeps its context in ThinkingMemory, and runs SQL, SSH, API calls, and MCP tools on its own. A heartbeat check watches your infrastructure and routes whatever breaks to Slack, Discord, Teams, email, or PagerDuty. You write the checks in TL or PySpark.

    6 min readAutonomous Agentread article →
  11. 09
    February 28, 2026
    Data Deserves Its Own Language

    A compiled, statically typed language where tables, streams, tensors, and models are types the compiler knows about, rather than libraries bolted on. One language in place of the Python, SQL, Spark, and YAML you currently glue together.

    5 min readOpen Sourceexplore ThinkingLanguage →
  12. 08
    February 21, 2026
    We Gave the Agent a Terminal

    One workspace with a terminal, a chat, and a codegen tab, where the agent holds real tool access. It explores the database itself, writes the ETL, and runs the PySpark. No copying output between windows.

    7 min readAgentic AIread article →
  13. 07
    January 21, 2026
    Agents That Get Better the Longer They Run

    The first version of our open-source memory layer: layered storage and reasoning-aware retrieval, so an agent stops solving the same problem twice. We rebuilt it around a single recall primitive five months later, in post 16.

    8 min readAI Memoryread article →
  14. 06
    January 16, 2026
    Tech News in 60 Words

    A reader that cuts AI, tech, startup, market, and sport stories down to 60 words. Built because keeping ten tabs open to find one fact is not reading.

    3 min readNews Platformread article →
  15. 05
    January 13, 2026 · v1.0
    Your Schema Changed and Nobody Told the Pipeline

    Silent data corruption usually starts with an upstream ALTER TABLE that nobody announced. Studio catches the drift, works out what changed, and proposes the fix before the bad rows land in the warehouse.

    8 min readData Engineeringread article →
  16. 04
    January 12, 2026
    The Kafka Connect Layer, Connector by Connector

    A walk through the Kafka Connect ecosystem inside Studio: the 16+ connectors we ship, how change data capture behaves in practice, and bi-directional flow for event-driven work.

    6 min readData Streamingread article →
  17. 03
    January 12, 2026
    We Run on a VPS, and It Cut the Bill by 95%

    Self-managed Linux boxes instead of the big clouds. 95% off the server bill, full control of the machines, and the operational work we took on in exchange.

    5 min readInfrastructureread article →
Try it yourself

Half of these analyses were built by an agent in Bonacci Studio

The case studies above are real pipelines: the agent wrote the code, ran it on Spark, and rendered the results. Describe yours and watch it ship.

Start free