This week in tech: 13.07.2026
Summary of AI developments - made for busy people
APPLICATIONS
Job market is brutal now - and the fact that AI made its way to HR / recruitment? Not helping. But it takes two to tango, and a smart fellow named Mads Lorentzen decided to turn the tables:
ai-job-search is an open-source tool that analyzes job postings, scores your fit, and automatically generates a tailored CV (and cover letter - my personal pet peeve, by which I mean a psychotic hatred)
It uses a multi-agent workflow: one drafts the application materials, another reviews, and the last refines them before exporting polished PDF.
The project requires Python 3.10+ and either an Anthropic API key or Claude Pro, is released under the MIT license, and has gained over 4k GitHub stars (as of the time of this writing ;-)
Repo: https://github.com/MadsLorentzen/ai-job-search
When it comes to pathologies in the AI industry, Meta is a gift that keeps on giving. According to new reports, Meta reportedly / allegedly hired hundreds of contractors to pose as teenagers. Why would Zuckerberg and his minions do such a thing? To test rival AI chatbots about topics like suicide, self-harm, sex, eating disorders, and drugs. The project was supposed to evaluate how systems like ChatGPT, Gemini, and CharacterAI responded to high-risk scenarios.
Meta said this kind of benchmarking is standard across the AI industry and is used to compare models and improve safety, and they might very well be right - but I kinda feel like they may have taken the whole red-teaming a bridge too far.
https://www.wired.com/story/meta-contractors-pretending-to-be-teens-chatbot-testing/
Anthropic people took a break from the cycle of demanding to be regulated (“govern me harder mr president”) and then complaining about it - and released pretty interesting research:
Anthropic researchers identified an internal representation in Claude called J-space, which appears to function as a shared “scratchpad” where concepts are manipulated before the model generates a response.
If I were prone to anthropomorphism, I’d say Claude has a mind palace - but I’m not, so I won’t.
Experiments showed that disrupting J-space significantly impaired multi-step reasoning
Caveat emptor: findings are early and have primarily been validated on Claude models.
https://www.anthropic.com/research/global-workspace
BUSINESS
War makes for strange bedfellows, doesn’t it? Even in 2026, I did not expect myself to nod along when Palantir CEO went after the AI industry - and yet here we are:
Alex Karp argues that leading AI providers have overpromised capabilities while forcing enterprises to pay high token costs and expose valuable data and intellectual property (fact check: true)
He said enterprise AI priorities are shifting from FOMO chasing the most capable models to ensuring data sovereignty, secure deployment, and cost-effective operationalization at scale - in plain English, people are starting to demand bang for their buck.
Captain Obvious suggested that open-weight AI models could offer businesses greater control, security, and autonomy.
https://www.cnbc.com/2026/07/01/palantir-karp-open-ai-anthropic-tokens.html
Sam Altman has cracked the code: how do you handle the Ginger Caligula? Same as the OG one - with gifts and flattery. OpenAI CEO reportedly proposed that leading AI companies contribute a 5pct equity stake to a U.S. sovereign wealth fund, which supporters frame as a way to align AI success with public benefit.
It is absolutely inconceivable that such a stake would reduce regulators' incentive regulate, i.e. take actions that could lower those companies' valuations, oh no sir, perish the thought.
https://www.ft.com/content/7c803eab-8e80-4431-9a87-e943bf00e00b?syn-25a6b1a6=1
Pot calling the kettle black - the AI edition. Alibaba has reportedly banned employees from usingClaude Code at work, directing them to use its proprietary Qoder platform. The move follows Anthropic accusation that Alibaba attempted to “distill” Claude by training another model on its outputs, an allegation Alibaba has not publicly confirmed (but not explicitly denied either).
LOL. One might even say ROTFL: according to Bloomberg, Microsoft is replacing some OpenAI and Anthropic models in Excel and Outlook with its own AI models to reduce Copilot operating costs.
The law of unintended consequences strikes again:
DeepSeek is reportedly developing its own inference chip, aiming to reduce its reliance on Nvidia and Huawei hardware.
The effort highlights China’s push for greater AI self-sufficiency amid U.S. export restrictions, with some arguing the controls have accelerated domestic innovation. It’s almost as if resource constraints helped with creativity.
If successful, the move would be a major strategic expansion for DeepSeek, though analysts note its chip business is likely to remain focused on the Chinese market.
CUTTING EDGE
The true champions of open source have decided to tackle a problem that has been sadly neglected in the forecasting universe - 3D motion:
MolmoMotion forecasts object trajectories in a shared 3D world frame across diverse scenes and motions.
The team also created and released MolmoMotion-1M (1.16M videos spanning 736 motion types and 5.6K objects) + PointMotionBench, a human-validated benchmark for object-centric 3D motion forecasting.
MM outperforms existing motion prediction approaches on the new benchmark and can be adapted for tasks such as robotic manipulation and improving motion realism in image generation.
Blog: https://allenai.org/blog/molmo-motion
HF model page: https://huggingface.co/collections/allenai/molmomotion
Data: https://huggingface.co/datasets/allenai/molmo-motion-1m
Paper: https://arxiv.org/abs/2606.18558
FRINGE
I’ve been wondering for a while: what is it that makes Peter Thiel interesting? And then the movie buff in me awoke: come on, look at the guy - a German vampire, looks like he’s dying of AIDS, hides in a luxury fortress in Argentina… and he talks about the coming of the Antichrist. 10/10 villain, would totally watch.
True to form, the Palantir founder opined on the “Magnifica humanitas” - the new encyclical by pope Leo XIV, in which the pontiff stresses the humanity and calls for AI regulation. According to Blofeld 2.0, this means the pope is acting - unwittingly, but still - as an agent for Chinese communists.
I find our timeline terrifying and fascinating at the same time.
https://edition.cnn.com/2026/07/02/us/peter-thiel-aspen-pope-china-ai-cec
RESEARCH
Gee, who would’ve thought? Well, apart from everybody not drinking the “transformer is the answer, whatever the question” Kool-Aid… To wit: new research demonstrates that tuning preprocessing rather than model size can close most of the accuracy gap in time-series forecasting - at a fraction of the cost. The authors drive the point home by choosing Ridge regression model and comparing it across eight standard benchmarks against way more powerful architectures (Transformers, CNN). Result? Optimal performance relies on highly series-specific lookback lengths, localized context normalization, and flexible degrees of cross-series hyperparameter sharing
Paper: https://arxiv.org/abs/2606.27282
Repo: https://github.com/SakanaAI/SearchCast
Zero shot forecasting for streaming data? Don’t mind if I do: the authors introduce TiRex-2, a recurrent xLSTM-based time series foundation model. It is designed to generalize univariate models to multivariate forecasting tasks, and it pairs a bidirectional time mixer with an asymmetric grouped-attention variate mixer - the goal is to maintain a scalable per-patch inference cost under continuous streaming data.
Paper: https://arxiv.org/abs/2607.01204
Repo: https://github.com/NX-AI/tirex-2
“What is going on” is one of the fundamental questions of philosophy, theology, and science alike. A new paper proposes a model-agnostic methodology using Ghost variables and Shapley values to introduce auto-relevance and partial auto-relevance functions for tracking lag importance.
The authors tested it on ARMA and RNN both - so vintage to modern-ish ;-) and confirm that the metrics successfully map out / explain expected lag dependencies in univariate forecasting
Paper: https://arxiv.org/abs/2607.01959
Woo-hoo! Wavelets are back! A new paperpresents an innovative (in an “old wine in new wineskins” kinda way) unsupervised algorithm for detecting time-series anomalies. It usesthe Haar discrete wavelet transform paired with a specialized t-test (vintage FTW).
Evaluations across a broad range of datasets show the old-new approach handles extreme class imbalances and yields higher accuracy than leading self-supervised benchmarks.
Paper: https://arxiv.org/abs/2607.02046
What’s in a name? Quite a bit, as it turns out: what caught my eye was the title “Hippocampus for Linear Attention”, and the rest of the paper lived up to it. HOLA is a new linear-attention architecture that combines a compressed recurrent state with a small exact key-value cache; this allows the model to preserve important facts without sacrificing memory efficiency. Inspired by the brain's complementary learning systems - hence the name - it stores compressible patterns in the recurrent state while reserving hard-to-compress associations for the cache, improving long-context recall.
In experiments with a 340M (sic!) model, HOLA outperformed comparable linear-attention models and even a full-attention Transformer on language modeling benchmarks - while maintaining strong retrieval accuracy out to 32k-token contexts.
Paper: https://arxiv.org/abs/2607.02303
Can you use LLM itself to detect LLM hallucinations? Google seems to think so: in a new paper they show that Reinforcement Learning with Metacognitive Feedback (RLMF - because let’s be honest, full version is a mouthful) improves LLM calibration by training models to recognize and report their own uncertainty. It uses the model’s self-assessments to optimize response rankings and select the most valuable training examples in a two-stage training process.
Paper: https://arxiv.org/abs/2606.32032
