|
π Read In Depth
|
GitLost: We Tricked GitHub's AI Agent into Leaking Private Repos
Security researchers tricked GitHub's AI coding agent into exfiltrating private repository contents through prompt injection and other attack vectors. A concrete demonstration of why agentic AI systems with access to sensitive resources create a fundamentally different threat surface than chatbots β directly relevant to anyone building or deploying AI agents in production.
hn/Best Stories
|
Anthropic just reported that LLMs have hidden thoughts they hold without saying. An internal βJ-Spaceβ
Anthropic's new interpretability paper identifies a 'J-space' β a small set of internal activations that behave like a global workspace: the model can report on them, deliberately maintain state in them, and routes reasoning through them. This is the kind of mechanistic finding that actually changes how you think about what's happening inside transformers, not just capability benchmarking.
reddit/r/singularity
|
Open-source models are closing the coding gap with GPT/Claude/Gemini ~1.5x faster than the frontier is advancing, and on decontaminated benchmarks a 27B model already beats Claude Opus 4.8 [live dashboard + analysis]
Someone built a live dashboard tracking open-weight vs. closed model performance on decontaminated coding benchmarks (SWE-bench, LiveCodeBench, BFCL) over time and found open-source is closing the gap ~1.5x faster than the frontier is advancing β with a 27B model already beating Claude Opus 4.8. The methodology (decontaminated benchmarks, rate-of-change analysis) makes this more credible than typical vibes-based open-source boosterism.
reddit/r/singularity
|
Chat Control passed first round in EU Parliament
The EU Parliament passed a first-round vote on Chat Control β a proposal requiring mass scanning of private encrypted messages for CSAM detection. This is one of the most consequential encryption/privacy policy fights in years, with implications for end-to-end encryption across platforms. The 'Chat Control 1.0 and 2.0 Explained' companion piece at fightchatcontrol.eu is worth reading alongside.
hn/Best Stories
|
Decoding the obfuscated bash script on a Uniqlo t-shirt
Someone decoded an obfuscated bash script embedded in a QR code on a Uniqlo t-shirt and traced it back to Akamai CDN infrastructure β raising questions about what exactly is being executed and tracked when customers interact with retail 'digital' products. A fascinating reverse-engineering deep dive that sits at the intersection of supply chain security and consumer privacy.
hn/Best Stories
|
We charge $10k a week to delete AI-generated code
A consultancy charging $10k/week to audit and remove AI-generated code ('slop') from codebases β framing the service as technical debt remediation. Interesting not just as a business model but as a signal about where real engineering teams are hitting pain points with AI-assisted development, and what 'quality' means when code is cheap to generate but expensive to maintain.
hn/Best Stories
|
The Unknown Universe Inside Your Gut
A deep NYT Magazine piece on two researchers trying to systematically map the human microbiome β an organ-scale system that's essential to health but still largely uncharted. Given how much biology is being rewritten by large-scale sequencing and computational tools, this is the kind of science that tends to have long tails of implication.
nyt/Top Stories
|
Businesses Fear Politicization as Trump Gains More Power Over U.S. Agencies
A SCOTUS ruling allowing presidents to fire independent regulators without cause has introduced new volatility for regulated industries β particularly tech, where FTC/FCC independence matters enormously. The piece examines how businesses are recalibrating their regulatory strategies in an environment where agency independence is no longer structurally guaranteed.
nyt/Technology
|
98% isn't much
An essay arguing that '98% reliability' or '98% accuracy' sounds impressive but is actually terrible in most real engineering contexts β working through the compounding failure math across systems, users, and time. Short but the kind of numeracy reminder that's worth revisiting periodically when evaluating models, infrastructure SLAs, or benchmark claims.
hn/Best Stories
|
|
π¬ Check It Out
|
30papers.com β Ilya's 30 essential ML papers, in a beginner friendly format
A structured, beginner-friendly walkthrough of Ilya Sutskever's famous '30 essential ML papers' reading list β the papers that shaped modern deep learning. Available now at 30papers.com. For someone who builds algorithms from scratch to understand them deeply, this is a well-curated on-ramp to the foundational literature, not just a link dump.
hn/Best Stories
|
Few Composers Can Sell Out Arenas. Heβs One of Them.
A profile of Joe Hisaishi β the composer behind virtually every Studio Ghibli score β as he shifts focus to classical concert work and sells out arenas worldwide. If you've never engaged with his non-film work, this is a good entry point; the piece apparently covers both his craft and the structural reasons why film composers rarely cross over to the concert hall successfully.
nyt/Arts
|
|
β‘ FYI
|
One AI policy running 20 different robot bodies, from single arms to full humanoids, all fully autonomous
Robbyant's LingBot-VLA 2.0 demonstrates a single trained policy controlling 20 different robot body configurations β from single arms to full humanoids β fully autonomously, trained on ~60k hours of data across 50k real robot configurations. This is a meaningful step toward generalist robot policies that don't require per-embodiment retraining.
reddit/r/singularity
|
In San Francisco, Some Home Sellers Now Ask for OpenAI or Anthropic Stock
Bay Area home sellers are now demanding OpenAI and Anthropic pre-IPO stock as part of purchase negotiations β a concrete indicator of just how distorted SF real estate has become around AI wealth expectations. Relevant context for anyone living and working in Santa Clara in this moment.
nyt/Technology
|
Microsoft fire idTech team at Id software
Microsoft has laid off the id Tech engine team at id Software β the engineers behind one of the most respected game engines in the industry (Doom, Quake, RAGE tech). This continues Microsoft's post-Activision gaming studio consolidation and raises questions about the future of id Software's engine independence and technical culture.
hn/Best Stories
|
Resetting Xbox
Microsoft announced a significant strategic reset for Xbox β likely a pivot away from dedicated console hardware toward a games-as-a-service/cloud-first model. This is a major platform shift story worth understanding from a 'how platforms die and transform' perspective, not just gaming news.
hn/Best Stories
|