AI & ML Research 3 Days

Last updated: May 17, 2026, 2:42 AM ET

Agentic AI & Coding Tooling

The enterprise push behind OpenAI's Codex deepened over the past 72 hours as Databricks integrated GPT-5.5 into agent workflows after the model set a new state of the art on the Office QA Pro benchmark, while Sea Limited's CPO revealed the company is deploying Codex across engineering teams to accelerate AI-native software development across Asia. The rollout is anchored by a secure Windows sandbox that gives Codex controlled file access and network restrictions, a prerequisite for enterprise deployment that OpenAI spent months hardening. Sales teams are already finding utility in the tool: Codex now generates pipeline briefs, meeting prep packets, and stalled-deal diagnoses from raw CRM inputs, automating work that previously consumed hours of manual summarization. Meanwhile, outside OpenAI's orbit, the CodeSpeak experiment over a 10K-plus-line repository showed that an AI-native workflow can absorb an entire codebase, though the author warned of compounding errors in dependency management that require human oversight.

LLM Evaluation & System Architecture

As agentic workflows proliferate, the evaluation problem is drawing urgent attention. A post on Recursive Language Models drew sharp lines between recursive architectures and earlier paradigms like ReAct and Code Act, arguing that self-looping models fundamentally change how agents reason through multi-step tasks. That structural shift makes traditional quality checks inadequate, which is why another author urged practitioners to build decision-grade scorecards instead of relying on "vibe checks". The need for rigor extends to inference itself: the next AI bottleneck is the inference system, not the model, the author argued, warning that latency and throughput constraints will cap what even state-of-the-art models can deliver in production. Together, these posts sketch a maturing field where architecture, evaluation, and infrastructure must advance in lockstep.

Financial Services & Credit Risk

Two posts this week zeroed in on finance as a proving ground for agentic AI. A practical guide to credit scoring walked through turning raw borrower data into risk classes using tree-based models and feature engineering pipelines, while MIT Technology Review examined data readiness for agentic AI in financial services, noting that banks operate under heavy regulation and must process market events updated by the second. OpenAI is betting on this sector too: ChatGPT Pro users in the U.S. can now link financial accounts for AI-powered insights grounded in personal spending, savings, and investment data, though the post did not specify whether third-party auditors have reviewed the data handling pipeline.

AI Safety, Sovereignty & Ethics

The tension between capability and control surfaced across multiple outlets. OpenAI and Malta announced a partnership to offer Chat GPT Plus and AI training to all citizens, framing the rollout as a means to build "practical AI skills and use AI responsibly," though critics will likely ask whether a single vendor's product meets that standard. On the sovereignty front, MIT Technology Review argued that enterprises traded data control for early capability gains when feeding proprietary information into third-party models, and that the reckoning is arriving as autonomous systems demand more persistent data access. The ethical stakes sharpened with a report on deepfake porn that used real people's likenesses, and with MIT Technology Review's look at how Chinese short dramas became AI content machines, where generative tools are mass-producing serialized video at scale with minimal human direction.

Practical ML & Self-Study

On the learning side, a 12-month self-study roadmap from data analyst to data engineer laid out specific tools and projects, while an investigation into why a coding assistant replied in Korean to a Chinese prompt offered a concrete case study of how code vocabulary reshapes embedding spaces across languages. For those already using Claude, two posts covered continual improvement techniques and strategies for writing robust code, both emphasizing prompt versioning and test-driven feedback loops as the backbone of reliable AI-assisted development.