This week in tech: 22.06.2026
Summary of AI developments - made for busy people
APPLICATIONS
Everybody knows LLM are trained on everything their creators could get their hands on - legally and otherwise - but what specifically goes into the stew? Say hello to ModSleuth: a tool that reconstructs the model and dataset lineage
Maps an LLM’s full “family tree” by analyzing papers, model/dataset cards, code configs, and upstream artifacts.
Uncovers deep dependency networks, with chains up to 8 hops deep.
Surfaces provenance issues, including hidden license inheritance, train/eval coupling, documentation inconsistencies, and models used as judges / filters / data generators.
Paper: https://arxiv.org/abs/2606.12385
Demo: https://modsleuth.cal-data-audit.org/
Google introduces Open Knowledge Format (OKF): an open standard for organizing organizational knowledge as structured Markdown files (heavily inspired by Karpathy’s LLMWiki and the like). Having a standard, as opposed to yet-another-service, handles the main challenge: connecting LLMs to fragmented company knowledge.
OKF addresses this by treating each concept, dataset, process, metric, or API as a separate Markdown file that can be linked into a machine-readable knowledge graph. The whole framework is vendor-neutral: Google is publishing the specification, examples, and visualization tools as an open standard.
Repo: https://github.com/GoogleCloudPlatform/knowledge-catalog/tree/main/okf
Outsourcing your thinking to an LLM is a bad idea - episode 89167: Starbucks Korea temporarily closed all stores for a mandatory employee history lesson after a promotion built using ChatGPT. The campaign reference symbols and phrases linked to South Korea’s past, specifically the 1987 torture death of student activist Park Jong-chul.
https://www.theguardian.com/world/2026/jun/16/starbucks-korea-shut-all-stores-tank-day-promotion
Dario Amodei is not the most stable person ever when it comes to visions of the future: depending on day of the week and the phase of the moon, he might tell you that AI will bring about the era of post-scarcity - or an economic apocalypse, where all human jobs will be wiped out (who will buy Anthropic models then, and with what, is not specified). If you were worried about the doomp**n take, I have some good news: at least one job is safe.
According to a new study, replacing human participants in A/B tests with LLMs is not a drop-in substitution. Randomized experiments identify treatment effects by design, while LLM-based estimates identify them only under statistical assumptions. TL;DR LLM surrogates can be useful for previously studied settings (known unknowns), but for genuinely novel treatments (unknown unknowns) their validity cannot be verified in advance => human experiments remain essential.
Paper: https://arxiv.org/abs/2606.17165
Crashing expensive tech on a budget is one of my favorite kinds of stories, and this week we have a new one: Tesla driver-monitoring systems works by detecting a human face near the front, which is interpreted as driver being attentive. Turns out a cheap plastic doll generates a false positive.
I obviously don’t know the inner working of the Tesla solution, but it sounds to me like a really poorly trained image detection model. For a company at the forefront of AI, that’s a bit embarrasing.
https://electrek.co/2026/06/15/chinese-drivers-plastic-heads-fool-tesla-autopilot-camera/
BUSINESS
Karma in AI (1): a German court ruled that Google is legally responsible for defamatory falsehoods generated by its AI Overviews, rejecting the argument that AI-generated content should be treated differently from publisher-created content.
Google defense was pretty much about having the cake and eating it too: they are a glorious AI company, but LLM outputs are inherently unreliable and should not be trusted for important matters (then why are you peddling them EVERYWHERE).
Karma in AI (2): Anthropic is facing a lawsuit alleging that its Claude Max subscription plans deliver significantly less usage than advertised, despite being marketed as offering 5x and 20x the limits of the Pro tier. I'm shocked: shocked to find that Amodei and his minions were stupid enough to get caught.
You cannot make this up:
Github is owned by Microsoft
Github cannot get enough Azure compute to keep the lights on
Microsoft asks Amazon for help and get them some AWS access
https://www.businessinsider.com/microsoft-github-amazon-ai-cloud-capacity-2026-6
Elon got himself a new toy: SpaceX is acquiring Cursor for USD 60B
Cursor will operate as a SpaceX subsidiary and gain access to xAI’s Colossus supercomputing infrastructure for future model development.
The mothership control a leading AI coding platform - a direct competitor with Anthropic’s Claude Code and OpenAI’s Codex
CUTTING EDGE
Fable is over (for now), but have no fear - open weights are here. Say hello to GLM 5.2:
Focused on long-horizon coding and engineering tasks: designed for autonomous software engineering, agent workflows, and complex multi-step problem solving
1 million-token context window
MIT license
Claims state-of-the-art coding performance, but independent validation of all benchmark claims is still limited.
Native in Transformers from day one
Release blog: https://z.ai/blog/glm-5.2
FRINGE
Dealing with other people’s bs is part of why we are paid for working, but every once in a while you have to wonder whether it’s worth it. One such recent example is an American judge who cancelled a trial after realizing BOTH sides were using AI to generate their filings (and not a very good one at that, because there was a ton of hallucinations).
https://futurism.com/artificial-intelligence/judge-cancels-trial-lawyers-both-sides-ai
RESEARCH
Having grown up with Sepultura, I was immediately drawn to the tile: a new research presents ChronosAD, a framework that adapts zero-shot time series foundation models for continuous anomaly detection. The core idea is evaluating data within non-overlapping contexts - the framework leverages pre-trained forecasting weights to track violations without requiring local re-training.
Paper: https://arxiv.org/abs/2606.01300
Less is more - time series edition. The authors propose a 7-million parameter foundation model optimized for real-time CPU deployment in industrial edge scenarios. It introduces a Shifted Attention mechanism to leverage exogenous covariates alongside a synthetic data generator to address training scarcity.
Paper: https://arxiv.org/abs/2606.10798
Sometimes all you have is a black-box API, and still need to adapt a foundation model. If it’s time series you’re dealing with, ORCA is here to help: framework that learns error patterns conditioned on both inputs and model outputs. It improves online forecasting performance without requiring access to model parameters.
Paper: https://arxiv.org/abs/2606.14222
I am intrigued the authors introduce Time-Series-to-Structured-Program (T2SP): a representation that converts time series into symbolic programs describing trends, periodicities, and events. This allows off-the-shelf LLMs to perform reasoning tasks - question answering, editing, and captioning. Initial evidence suggests it’s more efficient than relying purely on the numerical aspect.
Paper: https://arxiv.org/abs/2606.12481
Stationarity time series have a nice and elegant theory behind them - the only problem is real life series usually do not behave that way. This new research proposes a retrieval-augmented forecasting system: retrieval mechanism that accounts for regime shifts and changing temporal dynamics, improving the relevance of retrieved historical segments.
Paper: https://arxiv.org/abs/2606.04135
