Daily Briefing

April 6, 2026

2026-04-05

27 articles

LinkedIn is secretly scanning your browser for 6,000 extensions, and you weren't told

2026-04-05

Summary

LinkedIn's practice of scanning over 6,000 extensions in the Chrome browser and collecting device fingerprints without user consent has been exposed as 'BrowserGate'.

Key Points

LinkedIn quietly scans for the presence of 6,222 Chrome extensions upon visit via the 'Spectroscopy' system
Generates a device fingerprint by collecting 48 hardware and software characteristics such as CPU core count, memory, and screen resolution
The collected fingerprint is encrypted with an RSA public key ('apfcDfPK') and attached to all API request headers in the session
This practice is not listed in LinkedIn's privacy policy and was revealed through an investigation by the European group Fairlinked e.V.
The scan list includes over 200 of LinkedIn's competing sales tools like Apollo, Lusha, and ZoomInfo

Notable Quotes & Details

Notable Data / Quotes

2.7MB JavaScript bundle
6,222 concurrent requests
Collection of 48 device characteristics
Encryption key identifier: 'apfcDfPK'
Independent verification completed by BleepingComputer

Intended Audience

General readers, security researchers, those interested in privacy policy

Microsoft calls Copilot 'entertainment only' while charging $30 a month for it

2026-04-05

Summary

The contradiction of Microsoft charging $30 a month while specifying Copilot as 'for entertainment purposes only' in its terms of use is drawing attention, with actual paid users accounting for only 1/30 of the target.

Key Points

Microsoft Copilot's terms of use include a warning that it is 'for entertainment purposes only and should not be used for critical advice'
The terms were updated in October 2025 and became widely known in April 2026
Microsoft 365 Copilot for enterprise is excluded from this clause, which only applies to consumer products
Microsoft spent about $80 billion on AI-related capital expenditures in FY2025 and invested $13 billion in OpenAI
The adoption rate is low, with actual paid users numbering less than 1/30 of the target users

Notable Quotes & Details

Notable Data / Quotes

'Copilot is for entertainment purposes only.'
Microsoft 365 Copilot price: $30/user/month (enterprise), $18 (business)
FY2025 AI capital expenditure approx. $80 billion
OpenAI investment $13 billion
Paid adoption rate: Less than 1/30 of target users

Intended Audience

Corporate IT decision-makers, general readers

Recap: Europe's top funding rounds this week (30 March – 5 April)

2026-04-05

Summary

A weekly funding summary of major startup investment rounds in Europe from March 30 to April 5, 2026.

Key Points

Mistral AI: Raised $830 million in debt — for the purchase of 13,800 Nvidia chips for a data center in Bruyères-le-Châtel, south of Paris, scheduled to go live in Q2 2026
IQM Quantum Computers: Raised €50 million from BlackRock, pursuing a Nasdaq listing via SPAC merger (valuation approx. $1.8 billion, expected June 2026)
Midas (Berlin): $50 million Series A — a platform that tokenizes institutional investment strategies into on-chain products
Standing Ovation (Paris): €30 million Series B — a precision fermentation startup producing casein from dairy waste
Kestra (France): $25 million Series A — an open-source data, AI, and infrastructure orchestration platform, with enterprise revenue growing 25x over 18 months

Notable Quotes & Details

Notable Data / Quotes

Mistral AI $830M debt financing (13,800 Nvidia chips)
IQM €50M, valuation approx. $1.8B
Midas $50M Series A, over $1.7 billion in total assets issued
Kestra 25x enterprise revenue growth over 18 months, used by over 30,000 organizations worldwide

Intended Audience

Investors, startup stakeholders, tech industry personnel

In Japan, the robot isn't coming for your job; it's filling the one nobody wants

2026-04-05

Summary

Japan is actively adopting physical AI (robots) as a national survival strategy to solve problems of population decline and labor shortages.

Key Points

The Japanese Ministry of Economy, Trade and Industry announced a goal to build a domestic physical AI industry and capture 30% of the global market by 2040 (March 2026)
Japan holds a strong position, accounting for about 70% of the world market for industrial robots as of 2022
Japan's population has decreased for 14 consecutive years as of 2024, with the working-age population ratio at 59.6% and projected to decrease by about 15 million over the next 20 years
Key drivers for physical AI adoption: Cultural acceptance of robots, demographic labor shortages, and strengths in mechatronics and hardware supply chains
The US and China are progressing at a faster pace in developing integrated full-stack systems for hardware, software, and data

Notable Quotes & Details

Notable Data / Quotes

Target of 30% global physical AI market share by 2040
70% share of industrial robot world market (2022)
Working-age population at 59.6%, projected to decrease by 15 million over the next 20 years
'Physical AI is not a matter of mere efficiency, but of industrial survival' — Sho Yamanaka, Salesforce Ventures

Intended Audience

Industrial policy stakeholders, robotics/AI investors, tech industry personnel

I let Gemini in Google Maps plan my day and it went surprisingly well

2026-04-05

Summary

A review of using Gemini AI integrated into Google Maps, finding it more useful than expected for planning a day's schedule based on public transportation.

Key Points

Gemini in Google Maps ('Ask Maps') uses a text box for conversation and provides answers based on map data and user reviews
Can plan schedules reflecting specific conditions such as using public transport, the order of lunch, walk, and cafe, and return time
External information such as weather can be linked, and suggestions can be modified through conversation if the first proposal doesn't fit
The reporter received a satisfying schedule consisting of a taco restaurant (Tacos Chukis), a plant shop, and a Scandinavian-style cafe
Evaluated as effective for discovering hidden places that are not well-known

Notable Quotes & Details

Intended Audience

General consumers, Google Maps users

Notes: A personal experience review that may include some promotional content

Meet 'AutoAgent': The Open-Source Library That Lets an AI Engineer and Optimize Its Own Agent Harness Overnight

2026-04-05

Summary

AutoAgent, an open-source library that allows AI to engineer and optimize its own agent harness (system prompts, tools, orchestration), has been released.

Key Points

AutoAgent minimizes human intervention as a meta-agent runs a loop to automatically modify, experiment with, and improve the agent.py (harness file)
Achieved 1st place on SpreadsheetBench (96.5%) and 1st place for GPT-5 score on TerminalBench (55.1%) after a 24-hour run
Applies Andrej Karpathy's concept of autoresearch (automating ML training loops) to agent engineering
A human only writes instructions in program.md, and the meta-agent repeatedly modifies agent.py directly
Experiment history is automatically recorded in results.tsv, which the meta-agent uses to learn the direction for the next experiment

Notable Quotes & Details

Notable Data / Quotes

1st place score on SpreadsheetBench: 96.5%
1st place for GPT-5 score on TerminalBench: 55.1%
Developer: Kevin Gu (thirdlayer.inc)

Intended Audience

AI engineers, developers, researchers

Inside the Creative Artificial Intelligence (AI) Stack: Where Human Vision and Artificial Intelligence Meet to Design Future Fashion

2026-04-05

Summary

An article overviewing how AI is emerging as a core tool in the fashion industry across design, trend forecasting, and production.

Key Points

According to the McKinsey 2026 State of Fashion report, over 45% of global apparel brands have adopted AI-based design tools to shorten development lead times
Generative AI tools like Adobe Firefly and Midjourney are used for mood boards, sketches, and 3D prototyping
Tools like Fashion Diffusion automate visual tasks to shorten iteration cycles
Multimodal AI is used for micro-trend forecasting by simultaneously analyzing text, image, and video data
Large companies like WGSN perform trend forecasting 4-5 seasons ahead, and more brands are linking real-time customer feedback with design trends using AI

Notable Quotes & Details

Notable Data / Quotes

Over 45% of global apparel brands adopted AI design tools (McKinsey 2026 State of Fashion)
WGSN: Trend forecasting 4-5 seasons ahead

Intended Audience

Fashion industry personnel, AI researchers, design students

Notes: An overview-style article including educational advice for students

Show GN: Ravenclaw - An Open-Source System for Managing AI Coding Agent Task Context

2026-04-05

Summary

A post introducing Ravenclaw, an open-source system that can consistently maintain and reload task contexts across sessions among multiple AI coding agents like Claude Code, Gemini CLI, and Codex.

Key Points

Task context is accumulated in Ravenclaw regardless of which AI agent (Claude Code, Gemini CLI, Codex) is used, and previous situations can be reloaded as they were using MCP tools in a new session
Provides over 40 tools via the MCP protocol, with full features available through CLI and REST API
Epics/issues structure, graph views, and progress per project can be understood at a glance in the web UI
Supports a method where the agent sends a Human Input Request when judgment is needed, and the user answers in the web UI
Self-hosted, runnable with just PostgreSQL, Apache 2.0 license

Notable Quotes & Details

Notable Data / Quotes

GitHub: https://github.com/chainofdive/ravenclaw
Apache 2.0 license
Over 40 tools provided via MCP protocol

Intended Audience

Developers utilizing AI agents

Fastest Open Source Project to Hit 100k Stars in GitHub History (Sigrid Jin & Bellman)

2026-04-05

Summary

A post introducing the 'Oh My Codex' system and related tools that automate GitHub issue management, PR merges, and testing via AI agents, sharing the possibilities of AI agent-based development workflows.

Key Points

Oh My Codex automatically swarms multiple agents based on tmux sessions to autonomously handle GitHub issues, PRs, and tests
Built-in 'AI Slop Cleaner' skill to automatically clean up low-quality code (AI Slop) generated by AI
Advanced enough to complete project scaffolding via text commands even in poor Wi-Fi environments
Emphasizes the philosophy that humans should lead system design and take on the role of coordinating AI agents
Mentioned connection with claw-code (a Python clean-room rewrite project based on leaked Claude Code source)

Notable Quotes & Details

Notable Data / Quotes

YouTube video: https://www.youtube.com/watch?v=RpFh0Nc7RvA

Intended Audience

Developers utilizing AI agents

Notes: Incomplete content as it mixes YouTube Gemini summary content and GeekNews community reactions

Apple Approves Driver Allowing Nvidia eGPU Use on Arm Macs

2026-04-05

Summary

News that Apple's signature approval of a driver developed by Tiny Corp has opened the possibility of running LLMs on Arm Macs using Nvidia eGPUs.

Key Points

The driver was developed by Tiny Corp, not Nvidia, and can be used without disabling SIP (System Integrity Protection) thanks to Apple's signature approval
Requires direct compilation through Docker; not a typical plug-and-play method
Designed for running LLMs, with Apple approving both AMD and Nvidia drivers
Many community opinions suggest practicality is low due to Thunderbolt bandwidth limitations, and buying a used PC is better for LLM purposes
Currently only works exclusively for Tinygrad, and CUDA/Vulkan cannot be used in PyTorch

Notable Quotes & Details

Notable Data / Quotes

Apple has refused to sign Nvidia eGPU drivers since 2018
The Nvidia GPU network mount method within a LAN has about 4% overhead

Intended Audience

Mac users, ML/AI developers

Does Emotional Expression Change AI Performance? — Real Effects of Prompt Emotional Framing

2026-04-05

Summary

Introduction to research where Harvard researchers found through experiments with 6 benchmarks that fixed emotional prefixes have almost no impact on LLM performance, but adaptive emotion selection (EmotionRL) can lead to consistent performance improvements.

Key Points

Fixed emotional prefixes (e.g., 'I am asking because I am angry') do not significantly affect performance in most task-model combinations
Even when increasing emotional intensity (e.g., 'I am extremely afraid'), accuracy does not change proportionally to intensity
Adaptive emotion selection (EmotionRL) conditioned on input achieved performance improvements exceeding average static emotion baselines across 5 tasks
Evaluated in a zero-shot inference environment with three open-source models: Qwen3-14B, Llama 3.3-70B, and DeepSeek-V3.2
The researchers propose redefining emotional prompting as an 'adaptive routing problem' rather than a 'universal template'

Notable Quotes & Details

Notable Data / Quotes

arXiv:2604.02236v1
Authors: Minda Zhao, Yutong Yang, et al. (Joint research by Harvard and Bryn Mawr)
Used 6 basic emotions based on Plutchik's theory: happiness, sadness, fear, anger, disgust, surprise
Variance across models and emotions appeared most prominently in SocialIQA

Intended Audience

AI researchers, developers interested in prompt engineering

How Many Products Use Microsoft's 'Copilot' Name?

2026-04-05

Summary

A post analyzing and visualizing that the 'Copilot' name is used for over 75 products and features at Microsoft, showing a lack of consistency in the naming system.

Key Points

The 'Copilot' name is used for at least 75 products/features (including apps, platforms, keyboard keys, laptop categories, and Copilot creation tools)
Result of directly organizing information from various sources as there is no official complete list even within Microsoft
Produced an interactive visualization map grouping each Copilot by category and representing relationships with interconnecting lines
Perception in the community that 'Copilot' has effectively become a common name Microsoft attaches to LLM-based features
Google's 'Gemini', Apple's 'Apple Intelligence' are evaluated as similar strategies

Notable Quotes & Details

Notable Data / Quotes

At least 75 products use the 'Copilot' name
"In Linux, everything is a file; in Microsoft, everything is a Copilot" (Community reaction)

Intended Audience

IT industry personnel, developers, general readers

[D] Hash table aspects of ReLU neural networks

2026-04-05

Summary

A discussion post presenting a theoretical perspective that each layer of a ReLU neural network can be interpreted as a locality-sensitive hash table lookup and associative memory.

Key Points

If ReLU decision values are collected into a diagonal matrix D, the layer output is expressed as DWx
The product Wₙ₊₁Dₙ with the next layer weights can be interpreted as a locality-sensitive hash table lookup of a linear mapping
Can also be seen as an associative memory with Dₙ as the key
Related discussions are ongoing on the Numenta forum (gated-linear-associative-memory)

Notable Quotes & Details

Notable Data / Quotes

https://discourse.numenta.org/t/gated-linear-associative-memory/12300

Intended Audience

AI/ML researchers, developers interested in neural network theory

Notes: A brief preliminary level post in Reddit discussion format

Building behavioural response models of public figures using Brain scan data (Predict their next move using psychological modelling) [P]

2026-04-05

Summary

A post introducing the COGNEX system, which utilizes Meta's fMRI scan-based AI model TRIBE v2 to learn and predict behavioral response patterns of public figures.

Key Points

TRIBE v2, released by Meta, is a model that takes text, audio, image, and video as input and converts it into human brain processing methods
COGNEX uses TRIBE as a common reference frame to learn the response patterns of specific individuals and model deviations from the average baseline
Builds personalized behavior prediction models by collecting stimulus(event)-response(behavior/statement) pairs
Claims it can be utilized for information analysis, negotiation, and public message strategy
Open-source release planned, demo video released

Notable Quotes & Details

Notable Data / Quotes

TRIBE v2: Released by Meta about 2 weeks ago, trained on fMRI scan data
Demo: https://youtu.be/fVaTJXiJ9ZM

Intended Audience

AI researchers, information analysts, developers interested in psychological modeling

Notes: Promotional post with no mention of ethical considerations

Auto agent - Self improving domain expertise agent

2026-04-05

Summary

A post introducing AutoAgent, a self-improving agent system where a meta-agent repeatedly improves the harness (tools, system prompts) of existing agents to reach top performance in multiple domains within 24 hours.

Key Points

Starts from the perspective that the cause of agent performance degradation lies in the harness (tools, system prompts, etc.) rather than the model
A meta-agent automatically adjusts, tests, and repeatedly improves the harness, operating autonomously until the goal is reached
Using the same model (Claude) as an evaluator allows for efficient identification and improvement of failure causes
Achieved top rankings in two domains: Terminal benchmark (code) and spreadsheet (financial modeling)
Open-sourced on GitHub

Notable Quotes & Details

Notable Data / Quotes

Reached #1 in multiple domains within 24 hours
GitHub: https://github.com/kevinrgu/autoagent

Intended Audience

AI developers, agent system researchers

Notes: Promotional Reddit post

The person who replaces you probably won't be AI. It'll be someone from the next department over who learned to use it - opinion/discussion

2026-04-05

Summary

A Reddit discussion sharing opinions that role boundaries between strategy, product, and engineering jobs are blurring as people emerge who cross job boundaries using AI tools.

Key Points

Phenomenon where strategy personnel produce prototypes directly with Claude, engineers make product decisions, and product personnel implement strategy hypotheses directly
AI is dramatically increasing individual productivity, blurring boundaries between job functions
Currently a pattern mainly seen in the tech/big tech industry, with discussions on whether it will spread to other industries
The view that the entity taking jobs is not AI itself, but a colleague from another department who learned to use AI

Notable Quotes & Details

Intended Audience

Office workers, general readers interested in AI utilization

Gemma 4 26b is the perfect all around local model and I'm surprised how well it does.

2026-04-05

Summary

User experience sharing that Gemma 4 26B is the optimal model among local LLMs on a 64GB Mac, balancing speed, coding performance, and stability.

Key Points

Succeeded in implementing a Doom-style raycaster with just 3 prompts (Qwen series failed due to loops or rewrites)
Lower system load and faster response time compared to Qwen 3 Coder or Qwen 3.5
Characteristic of having concise thinking processes without being overly immersed in details
Expectation that local models could reach the level of Claude Sonnet within 2-3 years

Notable Quotes & Details

Notable Data / Quotes

Tested on a 64GB memory Mac
Implemented a Doom-style raycaster with 3 prompts

Intended Audience

Local LLM users, developers

Gemma 4 31B vs Gemma 4 26B-A4B vs Qwen 3.5 27B — 30-question blind eval with Claude Opus 4.6 as judge

2026-04-05

Summary

Sharing comparison results of a 30-question blind evaluation using Claude Opus 4.6 as judge, where Qwen 3.5 27B ranked 1st in number of wins but had a 10% failure rate risk, while Gemma 4 31B recorded the joint highest average score.

Key Points

Qwen 3.5 27B: 14 wins (46.7%), average score 8.17 (highest at 9.08 if excluding 3 zero points)
Gemma 4 31B: 12 wins (40.0%), average score 8.82
Gemma 4 26B-A4B (MoE): 4 wins (13.3%), average score 8.82, with errors in 2 questions
By category, Qwen was strong in reasoning/analysis, while Gemma 4 31B was dominant in communication
Claude Opus 4.6 judge had a 99.9% parse rate and used an absolute score (0-10) method

Notable Quotes & Details

Notable Data / Quotes

Total evaluation cost: $4.50
3 zero points occurred for Qwen 3.5 27B (CODE-001, REASON-004, ANALYSIS-017)
Claude Opus 4.6 parse rate 99.9%

Intended Audience

AI researchers, local LLM developers

Gemma 4 for 16 GB VRAM

2026-04-05

Summary

A user guide sharing the optimal quantization and parameter settings for the Gemma 4 26B A4B MoE model in a 16GB VRAM environment, and comparing performance against Qwen 3.5 27B.

Key Points

Recommended quantization: unsloth's gemma-4-26B-A4B-it-UD-IQ4_XS.gguf
Optimal parameters: --temp 0.3 --top-p 0.9 --min-p 0.1 --top-k 20
Recommended --image-min-tokens 300 --image-max-tokens 512 for improved vision performance
4x faster generation speed compared to Qwen 3.5 27B (80+ tps vs 20 tps), superior in multilingual, latest library utilization, and DevOps fields
Recommended to use llama.cpp b8660 build (tokenizer issues exist in the latest builds)

Notable Quotes & Details

Notable Data / Quotes

80+ tps vs 20 tps for Qwen 3.5 27B
Supports 30K+ token KV fp16
Tokenizer issues occurred in builds after llama.cpp b8660

Intended Audience

Local LLM users, AI developers

One year ago DeepSeek R1 was 25 times bigger than Gemma 4

2026-04-05

Summary

A brief opinion piece marveling at the speed of local LLM development by comparing DeepSeek R1, which was 671B parameters about a year ago, with the current Gemma 4 MoE at only 26B.

Key Points

DeepSeek R1: Released about a year ago, MoE structure, 671B parameters
Gemma 4 MoE: 26B parameters, approx. 25 times smaller than DeepSeek R1
Shows impressive performance despite the dramatic reduction in size
Expressing expectations for the rapid development of local LLMs

Notable Quotes & Details

Notable Data / Quotes

DeepSeek R1: 671B parameters (MoE)
Gemma 4 MoE: 26B parameters
Approx. 25x size difference

Intended Audience

Readers interested in AI/ML, local LLM users

Notes: A very short opinion piece

Comparing Qwen3.5 vs Gemma4 for Local Agentic Coding

2026-04-05

Summary

A post confirming that Qwen3.5-27B remains the best choice for reliability and code quality in local agentic coding tasks, based on a comparison with Gemma4 in a 24GB GPU (RTX 4090) environment.

Key Points

Qwen3.5-27B: Successful complex agentic coding in 1 attempt, highest code quality (using type hints, docstrings, pathlib), ~45 tok/s, 21GB VRAM, 130K context
Gemma4-26B-A4B (MoE): Fast at ~135 tok/s but lowest code quality, requiring retries
Gemma4-31B: Joint 1st with Qwen3.5-27B, code is clean but lacks depth, 65K context (limited by 4090)
MoE models have ~3x faster generation speed but require retries for complex tasks
All models showed a common failure in writing integration tests that call actual APIs despite TDD requests

Notable Quotes & Details

Notable Data / Quotes

Qwen3.5-27B ~45 tok/s vs Gemma4-26B-A4B ~135 tok/s
MoE model generation speed is about 3 times faster
Qwen3.5-35B-A3B: 32K token generation in complex tasks

Intended Audience

AI developers, local LLM users

Mercor Hacked: Major AI Companies' 'Training Secrets' at Risk After Supply Chain Attack

2026-04-05

Summary

AI training data provider Mercor was hacked via a supply chain attack exploiting the LiteLLM open-source library, putting training secrets of major AI companies like OpenAI, Anthropic, Meta, and Google at risk of exposure.

Key Points

Hacker group TeamPCP breached LiteLLM's CI/CD pipeline and distributed 2 types of malicious packages on PyPI (March 27), which were discovered and deleted after being deployed for about 40 minutes
The malicious code was designed to collect environment variables, API keys, SSH keys, AWS/GCP/Azure credentials, and Kubernetes configurations to exfiltrate them externally
Mercor confirmed a data breach of about 4TB, including 939GB of platform source code, a 211GB user database, and about 3TB of video interview recordings
Meta has indefinitely suspended its $10 billion (approx. 15 trillion won) collaboration with Mercor, and a class-action lawsuit has been filed targeting over 40,000 people
Security experts analyze this as a 'typical case of a sophisticated sequential supply chain attack' and warn that training methodologies themselves, not just datasets, may have been exposed

Notable Quotes & Details

Notable Data / Quotes

LiteLLM has approx. 97 million monthly downloads and is used in 36% of all cloud environments
Leaked data scale: 939GB source code + 211GB DB + approx. 3TB video files = total approx. 4TB
Meta's Mercor collaboration scale: $10 billion (approx. 15 trillion won)
Class-action targets: Over 40,000 former and current contract employees and customers

Intended Audience

AI security researchers, developers, AI company personnel

OpenClaw No Longer Supported on Claude Monthly Subscriptions due to 'Excessive System Burden'

2026-04-05

Summary

Anthropic has discontinued support for third-party tools like OpenClaw in its Claude subscription service to cope with surging demand.

Key Points

Boris Cherny, head of Claude Code, announced the discontinuation of support for third-party tools in Claude subscriptions starting the next day via X, drawing criticism for the one-day grace period
Third-party tool users must purchase 'additional usage packages (discount applied)' or use the Anthropic developer platform API separately
Anthropic explained that the subscription service was not designed for the usage patterns of third-party tools and also violates the terms of service
Claude's popularity has surged in recent weeks, reaching #1 in downloads on the US Apple App Store, and subscription usage limits were adjusted the previous week
Google similarly took action against third-party tool users of Gemini CLI based on terms of service violations

Notable Quotes & Details

Notable Data / Quotes

Cherny: "Our systems are optimized for specific workloads, and we are continuously optimizing to provide the most intelligent models to as many users as possible"
OpenClaw founder Peter Steinberger: "They dropped the bad news by surprise on a Friday night"
Claude achieved #1 in US Apple App Store downloads (last month)

Intended Audience

Claude subscribers, AI service users, developers

[Ahn Kwang-seop's AI Syntheses] What the Anthropic Source Code Leak Means

2026-04-05

Summary

A column critically analyzing the legal and ethical problems of derivative repositories claiming 'clean-room re-implementation' and Anthropic's internal double standards, following the Claude Code source code leak.

Key Points

On March 31, the entire source code of Claude Code (approx. 510,000 lines) was leaked and mirrored on GitHub due to a packaging error where source map files were included in the npm distribution version
When the original mirror and over 8,100 forks were blocked by Anthropic's DMCA takedown, one repository repackaged it as a 'Python clean-room re-implementation' based on the leak and transferred over 50,000 stars as they were
Clean-room reverse engineering must be re-implemented by a team that has never seen the original code based only on specification documents; re-implementing after directly reading the source code legally violates the premise of a clean room
Under the US Defend Trade Secrets Act (DTSA), use by those who knew or could have known that leaked information was acquired by chance or mistake is prohibited, and security vulnerabilities do not extinguish legal protection
The 'Undercover Mode' found in the leaked code reveals a double standard where Anthropic requires AI code attribution from others while creating a feature to erase AI traces for its own employees' open-source contributions

Notable Quotes & Details

Notable Data / Quotes

Leaked source code scale: approx. 510,000 lines
DMCA takedown targets: original mirror and over 8,100 forks
Repackaged repository stars: over 50,000 (transferred directly from the leak era)

Intended Audience

AI developers, legal/security stakeholders, open-source community

Notes: Author's personal column including legal interpretation; the author explicitly states they are not a legal expert

How I beat the $4 gas average in 2026: These 5 apps show you the cheapest station nearby

2026-04-05

Summary

An article introducing 5 mobile apps that find the cheapest nearby gas stations in a situation where gas prices have surged due to the war in Iran.

Key Points

In the aftermath of the war in Iran, average gas prices in the US rose to about $4 per gallon, with California nearing $6
GasBuddy (free on iOS/Android) displays a list of nearby gas stations, price per gallon, and user ratings based on location
Gas prices at nearby stations can also be checked in navigation apps like Google Maps and Waze
Each app provides filters for fuel types (regular, premium, diesel, etc.)
A recommendation article based on ZDNET's affiliate commission model, emphasizing adherence to editorial independence

Notable Quotes & Details

Notable Data / Quotes

Average US gas price: approx. $4 per gallon (up over $1 from the previous year)
Some areas in California: up to $6 per gallon

Intended Audience

General consumers, car drivers

Notes: Recommendation article based on ZDNET affiliate advertising. Low direct relevance to AI/tech keywords.

PreviousDaily Briefing