This week in tech: 6.07.2026
Summary of AI developments - made for busy people
APPLICATIONS
Ok, this is pretty awesome: tech saves lives
Google’s Android Earthquake Alerts system turns Android phones into a distributed seismic network using accelerometers to detect P-waves, estimate quake parameters in real time, and send warnings
During recent earthquakes in Venezuela - where no national early-warning system exists - it reportedly reached about 11M people, delivering up to 2 minutes of warning.
Early alerts likely enabled some people to take protective action, although exact casualty reductions remain unverified / speculative.
Mistral AI has released OCR 4 - new version of their document understanding model:
Combines high-accuracy OCR with structured layout analysis for large scale document processing.
Extracts text, equations, tables, charts, signatures, and document structure, returning bounding boxes and confidence scores for every detected element.
The model achieved an 85.2 score on OlmOCRBench, supports 170 languages,
OCR 4 is fully self-hosted, processes documents quickly at USD 2 / 1k pages in batch mode)
Announcement: https://mistral.ai/news/ocr-4/
On-device AI is the new black: Google introduced a new method for retrofitting Multi-Token Prediction onto existing Gemini Nano models, which accelerates on-device AI without retraining separate draft models. The approach addresses the efficiency bottlenecks of autoregressive text generation, delivering faster responses and lower energy consumption (= more bang for your buck). It is currently rolling out to Pixel 9 and 10 versions - I guess owning both the software and the hardware does have its perks.
Open-source benchmark IFStruct is here:
measures how well LLMs generate structured outputs from realistic instructions
IFStruct tests models on natural requests, annotated examples, JSON Schema, YAML, ASCII tables, mixed formats, and embedded constraints like “no code fence” or “no commentary.”
Outputs are scored with a binary pass/fail metric requiring every field, type, enum, bound, and count to be correct with no extra keys.
The dataset is generative, so it can also produce held-out training data, using the same validation check as a reward signal during training.
Strong results from small models, frontier models approach 100pct accuracy.
Blog: https://www.liquid.ai/blog/ifstruct-v1.0
Repo: https://github.com/Liquid4All/ifstruct
Meta has unveiled Brain2Qwerty v2:
- an open-source system that translates (non-invasive, allegedly) brain signals into text
- major step forward in brain-to-language interfaces. One the one hand, massive for people who are incommunicado for all sorts of reasons - on the other, the potential for things straight of Uncle Ted’s nightmares is unparalelled.
- Anyway: using MEG brain recordings from nine participants, the model achieved up to 78pct word accuracy and averaged 61pct
- Brain2Qwerty v2 predicts complete words and meaning
- Meta has released the full training code
Announcement: https://ai.meta.com/blog/brain2qwerty-brain-ai-human-communication/
Paper: https://ai.meta.com/research/publications/brain-to-text-decoding-a-non-invasive-approach-via-typing/
BUSINESS
Social media is linked to negative effects on children’s mental health / identity development / general well-being, so protecting youngsters is a good idea (even if some governments want to treat the problem by eliminating online anonymity - not a proportional reaction, to put it politely). Meta thinks we need to protect the corporations instead: they want a federal regulation to limit lawsuits alleging that Facebook and Instagram have harmed children through addictive platform design / algorithmic recommendations.
A bit of unsolicited advice: folks, if you want to be proper villains, then take a page from Peter Thiel’s playbook - foreign vampire, built a panopticon nightmare, hides in a fortress in Argentina, talks about antichrist all the time…
This? This is pathetic.
https://www.reuters.com/world/meta-lobbies-congress-protection-child-harm-lawsuits-2026-06-18/
First they came for Anthropic :
The US govt told OpenAI to delay the public release of GPT-5.6, with an initial rollout limited to a small group of approved enterprise partners following discussions with U.S. government officials.
It seems to be the new normal: model evaluation before broader public release due to national security concerns (what else duh).
It is ofc SOTA everywhere, with early results (partial release to trusted parties) showing that the model went all-in on benchmark hacking - for those with Kaggle in their past, this is optimizing for public leaderboard.
Chinese LLM are about to dominate the field, aren’t they.
While the EU is regulating everything from AI to AC, serious players seem to be inching closer towards a trade war:
U.S. expanded its blacklist of Chinese military-linked companies.
China imposed new trade restrictions on 56 U.S. companies, including export controls on 10 firms and procurement bans on 46 others
The measures target strategically important sectors such as rare earths and defense, affecting e.g. drone manufacturers
To the surprise of nobody outside the EU bubble, everything can be ued as geopolitical leverage - semiconductors, critical minerals, AI, you name it.
Anthropic has restored global access to Claude Fable 5 after U.S. export controls were lifted, adding a new cybersecurity classifier that “redirects high-risk requests to Opus 4.8” (gotta admire the openness: you pay for a SOTA model, but we will sometimes send your request to an older one - and no, we won’t tell you why).
My favorite part: Fable was not that special after all.
"Our testing confirmed that many less capable models—including Claude Opus 4.8, GPT-5.5, and Kimi K2.7—could identify the same vulnerabilities as Fable 5 did in the report. When it came to the demonstration of how to exploit the single vulnerability, every model we tested could produce the same demonstration as Fable 5 (including Claude Haiku 4.5, Sonnet 4.6, Opus 4.6, Opus 4.7, Opus 4.8, GPT-5.4, GPT-5.5, and Kimi K2.7)."
Anthropic has been working hard towards regulatory capture, so for them to get pwned by the Ginger Caligula? Fate does love irony.
https://www.anthropic.com/news/redeploying-fable-5
Meta is pivoting - again. Zuckerberg and his minions are reportedly building Meta Compute: a cloud business that would rent AI infrastructure and models, positioning the company to compete with AWS, Google Cloud, and Microsoft Azure.
Seems a bit late to the party, but hey - at least it makes more sense than the Metaverse.
https://techcrunch.com/2026/07/01/meta-like-spacex-looks-to-turn-excess-ai-compute-into-cash
Peace of mind has a price tag, and it’s apparently lower than expected: a Stanford study finds that >70pct of ChatGPT queries can already be answered accurately by local models. This suggests a large share of enterprise AI workloads could leave the expensive frontier API - local inference is becoming both accurate and fast enough to handle many real-world tasks. There is also a minor detail of dependency: if you run your own models locally, next time a Fable type of incident happens you can shrug and move on.
Paper: https://arxiv.org/abs/2511.07885
RESEARCH
Sometimes the most difficult course of action is to do nothing: it’s true in investing, in politics -and it’s true in machine learning. The authors of a new paper propose a selective forecasting framework that ‘permits| a model to abstain from making high-risk predictions on difficult samples. By leveraging meta-learning to model the empirical percentile of forecasting errors, the framework enables abstention transfer across heterogeneous time series.
Paper: https://arxiv.org/abs/2606.23448
Standard pointwise error metrics (like MSE) often overlook signal phase alignments and oscillatory behaviors - so the scary t-word has to enter the conversation. This paper proposes a topology-driven assessment tool: they combine delay embedding and persistent homology to extract intrinsic sequence dynamics. This allows for flagging of structural forecast degradations, which traditional metrics miss.
Paper: https://arxiv.org/abs/2606.25439
Producing coherent demand forecasts for supply chains with thousands of nested time series is complicated, expensive and (mathematically) restrictive. To scale efficiently, the authors introduce e2eTD method: the core idea is to isolate a tiny fraction of aggregate upper-level series for base forecasting - and then downsample the values through joint in-sample distribution modeling.
Paper: https://arxiv.org/abs/2606.26774
