This week in tech: 16.03.2026
Summary of AI developments - made for busy people
APPLICATIONS
Computer says yes - the Anthropic edition:
The U.S. military deployed Palantir’s Maven Smart System during strikes on Iran: it used AI to identify and prioritize up to 1000 targets within the first 24 hours of the operation.
Anthropic’s Claude was integrated into the platform to assist in generating coordinates (wtf), suggesting specific targets (wtf wtf), and performing post-strike reviews by processing data streams.
Due to “disagreements” with the Pentagon, Claude has been restricted from future government deployments and “will only be used temporarily” until a replacement system is finalized.
https://www.washingtonpost.com/technology/2026/03/04/anthropic-ai-iran-campaign/
Andrej Karpathy, a.k.a. The President of AI, open-sourced Autoresearch:
Tiny AI research system that autonomously runs ML experiments without supervision, letting an AI agent design and test improvements
The agent edits the training code, runs a 5-minute experiment, evaluates the result, keeps improvements, and repeats continuously on a single GPU
The entire project is three files (~630 lines) so the model can fit the whole codebase in context
The user sets research direction in
program.mdwhile the agent modifiestrain.py.
Repo: https://github.com/karpathy/autoresearch
Cloudflare CEO once positioned his outfit as defending content creators from heavy AI crawling and proposed a “pay-per-crawl” model - designed to help publishers monetize access. Alas, the times they are a-changin’: the company just launched a Crawl/Scrape API that simplifies large-scale site crawling and returns content in machine-friendly formats. It looks like Cloudflare is moving from restricting AI crawlers to commoditizing crawling infrastructure aligned with the demands of agentic AI systems.
https://developers.cloudflare.com/browser-rendering/rest-api/crawl-endpoint/
BUSINESS
Anthropic launched Claude Marketplace:
It allows enterprise customers to use their existing spending commitments to purchase third-party, Claude-powered software
Partners include GitLab, Snowflake, and Harvey
Anthropic centralized invoicing while taking no commission
CM simplifies procurement and allows companies to redirect budget toward a pre-approved ecosystem of tools.
The move incentivizes developers to build on Anthropic’s models while helping enterprises bypass the typical administrative, ekhem, hurdles of managing vendor relationships.
https://claude.com/platform/marketplace
At this stage I am beginning to think Zuckerberg and his minions want to eliminate humans from the loop completely: Meta just purchased Moltbook - the “social network for bots”.
https://www.axios.com/2026/03/10/meta-facebook-moltbook-agent-social-network
First Anthropic threw a tantrum when “trust us, we are the experts” did not fly with US Department of War, then they tried to grovel and bend the knee. When that did not work either, Anthropic remembered they are an American company after all - and switched to lawfare. Amodei and his minions are suing the administration after being labeled a “supply chain risk” - a designation that blocks it from military contracts. The company claims the decision came after it refused to allow its AI to be used for mass surveillance or autonomous weapons and argues the move is unlawful and politically motivated.
https://edition.cnn.com/2026/03/09/tech/anthropic-sues-pentagon
Google bought Wiz - a cloud and AI security company ran by a bunch of spooks (look up “Unit 8200”):
The company will be integrated into Google Cloud to create a “comprehensive security solution” for AI-driven enterprise environments.
Wiz will maintain its independent brand (for now anyway) and continue to support external platforms like AWS / Azure - the idea is to avoid (even a perception of) vendor lock-in.
Big business in bed with big government - getting into AI. What could possibly go wrong?
Nvidia and Palantir will build“AI operating system” - they want to control the underlying layer that dictates how all AI applications and data flows function.
The partnership seeks to replicate the historic platform-level power of Windows and Android: Nvidia’s hardware monopoly + Palantir’s expertise in, ekhem, let’s call it politely: converting massive data sets into government-level decisions.
If there is a civilization left a hundred years from now, future historians are likely to mark this incident as the beginning of the military-industrial-tech complex.
https://www.palantir.com/sovereignaios/
CUTTING EDGE
All your modalities are belong to us:
Gemini Embedding 2 simplifies multimodal search by unifying multiple embedding capabilities into a single model and API. This allows for cross-media search and retrieval without separate pipelines or preprocessing like transcription or frame extraction.
Default embeddings use 3072 dimensions with Matryoshka-style truncation support (= allowing developers to reduce vector size while preserving most semantic meaning)
It supports interleaved inputs (e.g., image + text in one embedding), handles long text and multiple media per request, and reportedly outperforms earlier Gemini embeddings and other multimodal models on retrieval tasks. No word on whether it cures common cold as well.
Announcement: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-embedding-2/
FRINGE
McKinsey’s AI chatbot Lilli has not had a good week:
It suffered a massive security breach after a basic SQL injection vulnerability was exploited
Decades of confidential strategy (LOL) and client data (decidedly not LOL) were exposed to an AI agent.
A research firm gained full read / write access to a production database containing millions of chat messages, confidential client records, and the system prompts that dictate the AI’s behavior.
The breach was triggered by unauthenticated API endpoints and a disturbingly simple architectural flaw: concatenating JSON keys into SQL - which allowed the attacker to bypass standard security scanners that typically only check data values.
The ability to rewrite system prompts means an attacker could silently manipulate the advice given to all the brilliant, hard working McKinsey consultants across the globe without triggering logs or versioning.
And these are the people in charge of digital transformation projects.
https://www.theregister.com/2026/03/09/mckinsey_ai_chatbot_hacked/
RESEARCH
“Artificial hivemind” is such an awesome phrase, I wish I had a prog metal band so we could use it as a title. In lieu of such masterpiece - I can’t play any instrument, sadly - I read a new paper from Stanford under this very title. The researchers have identified a phenomenon where different AI models, despite their (supposed) diversity, increasingly produce identical, homogenized responses to open-ended questions. This convergence occurs because most models rely on similar training data and alignment technique - which is a polite way of saying they all scraped the internet for the same data. As a result, all the models drift toward the same (predictable) patterns and phrasing.
Paper: https://arxiv.org/abs/2510.22954
This paper introduces the Amortized Predictability-aware Training Framework (APTF) to mitigate the instability caused by noisy (= low-predictability) samples. It uses a hierarchical loss function to dynamically identify and penalize such samples, and the adjustment improves performance in both forecasting and classification.
Paper: https://arxiv.org/abs/2602.16224
Speculative Speculative Decoding (SSD) improves LLM inference speed by parallelizing the drafting and verification steps. This allows the “draft” model to pre-compute multiple potential continuations while the target model is still verifying previous tokens. The predictions are stored in a "speculation cache" based on likely verification outcomes, which pretty much eliminates drafting latency. Tested on Llama-3.1-70B, this approach delivers up to a 2x speedup over standard speculative decoding and a 5x gain over autoregressive methods, particularly for low-batch-size workloads.
Paper: https://arxiv.org/abs/2603.03251
All foundation models use massive training corpora, and the time series variety is no exception. This paper warns that current benchmarks may suffer from significant "information leakage": it identifies overlaps in samples and temporal correlations that lead to overly optimistic performance estimates for time series foundation models.
Paper: https://arxiv.org/abs/2510.13654
The authors propose the Global Temporal Retriever (GTR), a plug-and-play module that extends a model's awareness to long-term periodic patterns beyond the immediate input window. It uses an adaptive embedding to align relevant global segments with the current sequence without the high costs of long historical windows.
Paper: https://arxiv.org/abs/2602.10847
