|
📖 Read In Depth
|
NSA says Mythos broke into almost all of their classified systems in hours, per The Economist
Anthropic's Mythos model reportedly broke into nearly all of the NSA's classified systems within hours, per The Economist. This is an extraordinary capability threshold — and the article context involves Trump's administration blocking Anthropic, adding a policy dimension. The emergent hacking capability (not explicitly trained) has serious implications for AI safety timelines and the whole 'alignment before capability' debate.
reddit/r/singularity
|
Mythos was not trained on 'hacking'. Other Ai labs also will reach Mythos-level capabilities in the future
Anthropic confirms Mythos wasn't specifically trained on hacking — the capability emerged from better coding, reasoning, and autonomy. This means other frontier labs (OpenAI, Google, even open-source) will likely reach similar capability levels soon. A significant data point for anyone thinking about AI risk timelines and what 'dangerous capability thresholds' actually mean in practice.
reddit/r/singularity
|
Anthropic built its name on AI safety. Can those commitments survive a trillion-dollar IPO?
Examines whether Anthropic's safety commitments can survive a trillion-dollar IPO — the classic tension between mission and capital pressure. Especially interesting given the Mythos news: the company that just demonstrated the most dangerous known AI capability is also the one most publicly committed to safety. The incentive structure here is worth thinking through carefully.
reddit/r/singularity
|
Colon cancer’s invisibility cloak removed by eliminating a single gene - a fundamental breakthrough. The result was 100% eradication of tumours when paired with immunotherapy treatment in mouse models.
Researchers achieved 100% tumor eradication in mouse models by knocking out a single gene that colon cancer uses to hide from the immune system, then pairing with immunotherapy. 'Fundamental breakthrough' is often oversold, but the mechanism here — removing an immune evasion cloak rather than attacking the tumor directly — is a genuinely different approach worth understanding.
reddit/r/singularity
|
Hey, n00b, we didn't hire you to complete tasks
Kent Beck argues that junior engineers are hired not to close tickets but to grow into people who can think about systems, tradeoffs, and the work that isn't yet defined. A sharp provocation about how task-completion framing stunts engineering development — especially relevant when AI handles an increasing share of the defined tasks, making the undefined ones more valuable.
hn/Best Stories
|
Elon Musk’s Next Move May Be a Mega-Merger of SpaceX and Tesla
Legal experts say a SpaceX-Tesla mega-merger is possible with little shareholder recourse — Musk's control structures at both companies give him unusual latitude. This would create a vertically integrated energy-transport-space conglomerate with implications for how compute, power, and capital flow across Musk's empire, including xAI.
nyt/Technology
|
Temporary Cloudflare accounts for AI agents
Cloudflare is introducing ephemeral accounts for AI agents — disposable identities that can make API calls, access services, and expire without leaving persistent credentials. This is infrastructure-level thinking about the agentic web: how do you give agents the access they need without the security footprint of persistent accounts? A concrete step toward agent-native architecture.
hn/Best Stories
|
|
🎬 Check It Out
|
Loupe – A iOS app that raises awareness about what native apps can see
Loupe is an iOS app that demonstrates what native apps can actually see on your device — camera, microphone, clipboard, location, and more — in real time. A practical privacy/security awareness tool that makes abstract data access concrete. Available on GitHub now; worth running just to see what permissions you've actually granted.
hn/Best Stories
|
|
⚡ FYI
|
Estimated share of newly written code that was AI-generated or AI-assisted
Aggregated data showing AI-generated/assisted code went from 6% in 2023 to ~42% currently, with 65% projected by 2027. Multiple sources (GitHub, Sonar) converging on similar numbers. The rate of change is the real signal here — this is a structural shift in software engineering happening faster than most org strategies can track.
reddit/r/singularity
|
Five Chinese AI Labs Cut Token Prices Up to 99%
ByteDance, Tencent, MiniMax, Alibaba, and Xiaomi all slashed token prices 50–99% in the same competitive window. BofA analysts attribute it to narrowing capability gaps between Chinese models. Classic commoditization dynamic — when differentiation collapses, price is the only lever, which accelerates infrastructure moats over model moats.
reddit/r/singularity
|
Nobel Winner John Jumper to Leave Google DeepMind for Anthropic
John Jumper, AlphaFold's lead and a 2024 Nobel laureate in Chemistry, is leaving Google DeepMind for Anthropic. Talent flow at this level is a strong signal about where serious researchers think the most important work is happening — and it will further concentrate biology/science-adjacent AI work at Anthropic.
reddit/r/singularity
|
Linux eliminates the strncpy API after six years of work, 360 patches
Linux 7.2 finally eliminates the strncpy API after six years, 360 patches, and systematic replacement across the entire kernel codebase. This is a meaningful safety win — strncpy's failure to null-terminate was a perennial source of subtle bugs. A good example of how long-term, unglamorous engineering hygiene actually works at scale.
hn/Best Stories
|
Reuters: Google to challenge German ruling saying it is liable for AI-generated false claims
Germany ruled Google liable for AI-generated false claims; Google is appealing. This is a landmark liability case that will shape how AI outputs are legally treated across the EU. The outcome could force architectural changes in how search-integrated AI products handle factual claims.
reddit/r/singularity
|
Hyundai takes full control of Boston Dynamics as SoftBank exits for $325 million
Hyundai bought out SoftBank's remaining stake in Boston Dynamics for $325M, taking full control. This consolidates the robotics company under a single industrial parent with manufacturing scale and a clear commercial roadmap. SoftBank exiting at this price — well below its earlier valuation — tells a story about how patient capital is running thin on long-horizon robotics bets.
reddit/r/singularity
|
D.O.J. Seeks to Halt Pollution Lawsuit Against Elon Musk’s Data Center
The DOJ is trying to kill an environmental lawsuit against xAI's data center in Mississippi, citing national security and xAI's role in the Iran war. This is a significant precedent: using national security to shield AI infrastructure from environmental accountability. The intersection of AI compute, pollution, and executive power is worth tracking.
nyt/Technology
|