This week in tech: 29.06.2026
Summary of AI developments - made for busy people
APPLICATIONS
A study of 26k Chinese students found that AI use improved homework scores (good) and reduced time spent on assignments (good) , but generally led to worse performance on closed-book exams (BAD). The decline was concentrated among students who used AI to complete homework much faster than their peers. This suggests that many were outsourcing, not using it for assistance - a conclusion supported by the data, because students who used AI but spent similar time on homework as non-users performed just as well on exams.
Paper: https://cepr.org/publications/dp21577
Health (1): Like it or not, ChatGPT has become a major health tool: > 230 million people use it weekly, and it’s apparently more than a glorified search. A study using OpenAI’s models helped doctors uncover 18 previously missed rare-disease diagnoses by generating evidence-backed leads for clinical review.
https://openai.com/index/improving-health-intelligence-in-chatgpt/
Health (2): Midjourney has had its moment as the undisputed king of image generation models, but their rule has come to an end and it looks the company has decided to reinvent itself. They just announced a full-body ultrasound scanner that can scan a person in about a minute and aims to deliver near-MRI-level detail.
The company also plans “Midjourney Spa” locations combining the scanners with wellness facilities (with a long-term goal of supporting large-scale medical diagnostics).
If they pull it off, it will be one hell of a pivot.
https://www.midjourney.com/medical/blogpost
Intermediate token generation is a fancy way of saying a model produces output before the solution, and it has become a standard method to improve the performance of language models on reasoning tasks. These intermediate tokens are routinely called reasoning - which implicitly anthropomorphizes them, and implies that these traces resemble steps a human might take (which, to be fair, is not completely baseless: not because models are smart, because a lot of humans are not).
In a new paper, the authors argue this attributing of human characteristics to a token prediction machine is a seriously bad idea - starting with, but not limited to, proliferation of questionable research.
Paper: https://arxiv.org/abs/2504.09762
OpenAI is stepping into the void created by Anthropic (what with Mythos being blocked and whatnot)
Daybreak is Altman’s cybersecurity initiative: an AI toolkit focused on helping organizations find, validate, and fix software vulnerabilities faster than attackers can exploit them.
The (upgraded) Codex Security plugin can scan codebases, trace attack paths, verify vulnerabilities, and generate patches
OpenAI fully released GPT-5.5-Cyber - the model ranks 86pct on CyberGym
There is also the outreach: expanded cybersecurity partnerships with the EU, among others, s to help protect critical infrastructure.
https://openai.com/index/daybreak-securing-the-world/
Less is more - image generation edition:
Say hello to i1: a fully open 3B-parameter text-to-image diffusion model
Released withthe data, code, and training recipe needed to reproduce / extend the work => ACTUAL OPEN SOURCE
Built entirely on publicly available datasets
Competitive with leading closed and open models alike across five major image-generation benchmarks.
Paper: https://arxiv.org/abs/2606.11289
Repo: https://github.com/zlab-princeton/i1
BUSINESS
Turns out OpenAI was not joking when they said they were going to build their own chip: together with Broadcom, they just presented Jalapeño - an LLM-optimized inference chip. It is scheduled for release 9 months, with OpenAI's own models being used to speed up the flow. Vertical integration FTW.
https://openai.com/index/openai-broadcom-jalapeno-inference-chip/
Anthropic is starting to repeat themselves: they pulled the omg-terminator-will-eat-us with GPT-2 and now with Mythos (and boy, did this one backfire), and now that their position at the top is threatened, they are accusing the evil Chinamen of stealing their stuff. Again.
To wit:
Anthropic alleges that the usual suspects among Chinese AI companies used geo-block circumvention, thousands of accounts, and large-scale prompting to extract Claude outputs for model training via a suspected “distillation” campaign. If that’s true, Anthropic defenses clearly have space for improvement.
The activity allegedly involved 25k accounts and 30M prompts over a 45-day period
The whole thing has an amusing karmic vibe to it: Anthropic built their models by stealing everything that wasn’t nailed to the floor (and sometimes the nails too), but when their stuff gets pinched? Oooh, property rights are suddenly the fundament of progress.
CUTTING EDGE
l
Announcement: https://www.liquid.ai/blog/lfm2-5-230m
HF model page: https://huggingface.co/LiquidAI/LFM2.5-230M
FRINGE
DuckDuckGo has introduced AI summary and this week it communicated to the users that Donald Trump has died of rabies. I was going to write something snarky about growing pains, but honestly: looking at the Middle Eastern policy of this administration (especially qua Iran and Palestine), I think the DDG LLM might be onto something here.
https://futurism.com/artificial-intelligence/duckduckgo-ai-trump-rabies
RESEARCH
MIT researchers introduced a variable-width Transformer architecture: they keeps early and late layers wide while narrowing the middle layers (the vintage one keeps the same width throughout the network). The challenger variant outperforms the incumbent on language modeling and most downstream benchmarks - and uses fewer resources (total FLOP reduced by 20pct, KV cache memory by 15pct).
Paper: https://arxiv.org/abs/2606.18246
Time series modeling suffers from the same problem as LLMs do: models are ranked on held-out average errors - or, as people half my age call it, benchmaxxing. The authors propose a benchmark named TS-Fault: it evaluates time series foundation models against four distinct real-world fault modes, detailing degradation levels across structural perturbations and persistent regime changes.
Paper: https://arxiv.org/abs/2606.18539
A new papr challenges the widespread implicit assumption that the structure of anomalies in multivariate time series is spread across cross-channel correlations. By evaluating eight public benchmarks with a diagnostic framework, the authors demonstrate that (almost) no cross-channel anomaly occurs without a simultaneous univariate deviation - which means channel-dependent deep models do not deliver much bang for the buck.
Paper: https://arxiv.org/abs/2606.02670
“What if” is one of the most commonly asked question in practical applications, and the if you want to approach it seriously, at some point you run into the problem of counterfactual generation (spoiler alert: tricky). The authors of this one reformulate counterfactual generation for time series as learning a globally consistent intervention strategy and not an instance-based optimization. The resulting architecture (ConTex) delivers targeted interventions across both temporal and feature dimensions, while lowering computational overhead.
Paper: https://arxiv.org/abs/2606.18049
You know how sometimes you build a time series model, the metrics look great, error has shrunk - but looking at the forecast graphs, something feels off? The core idea behind TimeVista is to formalize this hunch-driven approach to validation: the paper explores the use of Vision-Language Models (VLMs) as judges to evaluate the visual plots of time series forecasts.
