I attended a regional quasi-academic AI and Data Management day conference. I say "quasi-" because the panels were ranging in thought, but the in-between bits were fishbowl sales pitches to a captive audience [we've all been there - prop's to Google Cloud, they gave away socks which was novel). Here are my forward-looking takeaways from a professional practice point of view. The statements are drawn from my notes. Mileage may vary for my recollection is not perfect.
-
- The major platforms will not deploy armies of AI and ML consultants designed and intended to drive the use to their platforms. Reflexively, big tech will create robust platforms and AI/ML/Data ecosystems that foster pools of AI specialists outside their platforms.
- Max Tegmark (MIT) quote "The power of AI and ML technology cannot grow at a rate that exceeds the growth of the wisdom with which we manage it." Advances demand that we push forward into the AI and ML GRC workspace with equal zeal as the purist technology pursuits. In 2024, Cyber and GRC professionals will have to respond to a rapidly advancing modularization and decentralization AI ecosystem. A key hard problem in AI is, "Is AI governance scalable?"
- All AI and ML implementations fall into 4 generic patterns (a.k.a. thematic use cases): as an Assistant (helpful), as a Guardian (optimizing), as an Entertainer, and as an Oppressor (regrettably).
- An interesting researcher take on the impact of LLM's on software engineering productivity - allegedly its a three-layer pyramid. LLM's drive the productivity and optimize the work output of DEV experts, and entry-level staff using LLM's drive DEV education and training. The middle layer where most DEV's reside does not see much of a boost in productivity or learning. A interesting claim requiring more validation given that the conclusion has a strong conventional wisdom feel.
- Adopters of AI and ML capability often hit pilot fatigue not as a result of the tech introduction, but due to a host of data issues. Panelists cited weak data governance, dirty data, rapid loss of data quality while simultaneous drowning in poisoned (biased) data. Collectively this was referred to as "legacy data debt" similar in likeness to "technical debt."
- The culprit in data quality issues is a lack of data literacy on behalf of those who collect data for use.
- "Responsible AI" as a term is too narrow and already historically biased. The forward-looking term is "AI Safety" which is a combination of Data Governance and Privacy, Security and Compliance, Reliability and Sustainability, and Responsible AI.
- RAND data scientists very active in AI Futures choose a 2-year horizon due to the rate of change. The take-away is that typical 3-year IT planning is pointless in this space. The panelists stated that to track emergence, watch the calculated lag (time) between the release of proprietary frontier model capability and its closed open source counterpart. For example, GPT 3.5 LLM capability was 6 months ahead of Mistral OSS LLM models. The gap has now closed to 3 months in less than a year.
- Models are embedding into the hardware of IoT and IIoT products. Although not as splashy as GPT transformer architecture, AI and ML models are entering and optimizing the performance of smaller systems as embedded capability.
- If Neural Networks (NN) are fundamentally a curve-fitting exercise constantly adjusting weights in the intermediate layers of a long-input/output chain, then it stands to reason that the same NN's can identify biases in the data and correct for bias presence. The core problem is that the state of AI today is essentially a curve fitting exercise to a cloud of data points. Current state AI has no native capability to deal with counter-factual questions such as "what happens if this data set is biased for subgroups?" Which is a design objective and value alignment - Presently, AI only achieves objectives not design them.
- Going forward the only race that matters is AI Safety - a blend of Data Governance and Privacy, Security and Compliance, Reliability and Sustainability, and Responsible AI. AI Safety is (AI principles + Transparent ModelOps and MLOps Engineering + an AI Risk Framework)
- The emerging area of AI governance involves some combination of Model Cards (Google), Data Sheets for Data Sets (Microsoft), and "AI Nutritional Facts Labels (Twilio)" to explain what is "in the black box."
Mark Y. Co-Chair CSA AI Tech & Risk Workgroup.
------------------------------
Mark Yanalitis
------------------------------