HackerNews Digest

None

Kimi K3, an open‑source frontier model, was benchmarked against the closed‑source Fable 5 on ~1,030 real‑world agentic tasks spanning five task families. Using an “oracle router” that selects the cheapest correct model per task, K3 was chosen for 72‑96 % of tasks, yielding an overall 93 % accuracy when routing between the two models. Accuracy on individual benchmarks is close (e.g., SWE: K3 92.4 % vs. Fable 92.6 %); however, each model excels on different sub‑domains—K3 on symbolic math, dev tooling, terminal/cryptanalysis tasks, and legal work, while Fable leads on web/data visualization and multi‑language coding (Java, Python, C++). Cost analysis shows K3 can be up to 50× cheaper, mainly due to token pricing, prompt caching, and lower token usage on long‑horizon tasks. The study concludes that per‑task routing—using K3 as the default and falling back to Fable when optimal—outperforms any single model in both quality and cost efficiency.

LG Electronics announced it will suspend any webOS smart‑TV apps that embed residential‑proxy SDKs, giving developers a deadline to remove such functionality or face app suspension. The decision follows research by security firm Spur, which found that over 42 % of LG webOS apps and more than 25 % of Samsung Tizen apps contain SDKs that turn the TV into an always‑on proxy node. Most of these SDKs were linked to Bright Data, a proxy provider that compensates developers for inclusion. The SDKs appear in various apps, from games (e.g., Pac‑Man) to screensavers and utilities, often offering users a choice between ads and proxy participation via a one‑time consent prompt. LG’s senior vice‑president John Taylor said the company is reviewing its app store, strengthening evaluation processes, and will remove residential‑proxy options from future submissions. Spur emphasized the lack of user transparency and the risk of minors granting consent, while noting that proxy providers claim to enforce KYC and network‑level safeguards.

The provided excerpt contains only the title “Free Ink – An open ecosystem for e-readers,” indicating that the material pertains to a platform or framework named Free Ink designed to operate as an open, interoperable ecosystem for electronic reading devices. No further details, technical specifications, or contextual information are present in the supplied text.

The Jacobian conjecture asserts that a polynomial map F: ℂⁿ→ℂⁿ with constant non‑zero Jacobian is globally invertible. A recent counterexample in dimension 3 shows this fails. The construction begins with the multiplication map sending a linear homogeneous polynomial L and a quadratic homogeneous polynomial Q in two variables to their product C=L·Q, a cubic polynomial. This map, from a 5‑dimensional domain to a 4‑dimensional target, has scaling and GL₂(ℂ) symmetries, making it generically three‑to‑one and thus not globally injective. By fixing the resultant Res(L,Q)=1, the scaling symmetry is removed, yielding a four‑dimensional domain on which the map remains locally injective (the non‑vanishing resultant guarantees distinct roots) but still not globally injective. Choosing a three‑dimensional affine slice H in the target that avoids the origin (corresponding to a third‑order differential operator with a double root) produces a variety X that is birationally equivalent to ℂ³. Explicit coordinate calculations show that X is in fact polynomially isomorphic to affine space, while the restricted map on X is locally injective yet fails to be globally injective. Consequently the induced polynomial map ℂ³→ℂ³ has constant non‑zero Jacobian but no polynomial inverse, providing a concrete counterexample to the Jacobian conjecture in three dimensions.

None

The announcement introduces three new Gemini models aimed at higher token efficiency, lower latency, and reliable performance for production AI agents. - **Gemini 3.6 Flash** is positioned as the workhorse model, improving coding, knowledge‑work, and multimodal tasks. It reduces output‑token usage by 17 % versus 3.5 Flash (Artificial Analysis Index) and achieves up to 65 % gains on benchmarks such as DeepSWE (Datacurve). Pricing is $1.50 per 1 M input tokens and $7.50 per 1 M output tokens, lowering overall task cost. - **Gemini 3.5 Flash‑Lite** is the fastest, most cost‑effective 3.5‑class variant, delivering 350 output tokens per second (Artificial Analysis Index) and outperforming previous Flash‑Lite generations in agentic workflows. - **Gemini 3.5 Flash Cyber** pairs a specialized cyber‑focused model with the CodeMender security agent to provide competitive performance for cybersecurity applications. Future plans include broader release of Gemini 3.5 Pro (currently in partner testing) and the next‑generation Gemini 4, for which the most ambitious pre‑training run has begun.

The article outlines ten practical actions intended to increase personal happiness and improve social interactions. 1) Enhance sensory experiences through nutritious food, appealing photography, music, comfortable clothing, and mindful walking. 2) Pursue activities outside one’s competence (e.g., juggling, breath‑holding, Rubik’s cubes, music, drawing) to stimulate learning. 3) Increase contact with others by initiating conversations, smiling, and relocating to denser communities or adopting walking/cycling. 4) Join or create groups and projects, online or offline, to contribute to larger endeavors. 5) Record and regularly review completed tasks to reinforce a sense of achievement. 6) Identify and remove individuals who cause distress, referencing literature such as “The Psychopath Code.” 7) Recognize and manage negative emotions (anger, jealousy, fear, etc.) while reducing reliance on substances like alcohol. 8) Prioritize time for meaningful activities and relationships, even at the cost of reduced income. 9) Reduce material possessions to those that provide genuine satisfaction. 10) Adopt a non‑desiring, accepting stance toward experiences to limit disappointment.

A “canvas arena” was built that gives vision models a blank white canvas and a colored‑pencil toolset (set_color, set_brush, set_pressure, draw, smudge, erase, view_canvas, view_target, plan). Four models—GPT‑5.6 Sol, Claude Fable 5, Grok 4.5, and Gemini 3.6 Flash—were asked to (1) reproduce two reference images (Mona Lisa, Starry Night) and (2) generate drawings from five text prompts, for a total of 28 drawings. All runs used the same toolset; GPT‑5.6 and Gemini rarely called set_* tools, Grok used them heavily, and Claude relied on frequent smudge and self‑review calls. Similarity was measured by SSIM (0–1) on 256×256 resized images. Gemini 3.6 Flash achieved the highest raw SSIM on both targets (peak 0.449 on Mona Lisa) but its scores declined after mid‑run. GPT‑5.6 Sol showed consistent detail, peaking at 0.352 (Mona Lisa) and producing the most favored open‑prompt outputs. Claude Fable 5 was the most expensive (~$160 for seven drawings, ~20× GPT‑5.6) and slower, with limited quality gains. Grok 4.5 generated largely unusable results despite many tool calls. Token usage was highest for Grok (≈34 M) and Gemini (≈27.7 M) but remained cheap due to cached reads. The experiment demonstrates divergent tool‑use strategies, cost‑performance trade‑offs, and limited improvement after early plateaus in iterative self‑review.

A federal judge approved a $1.5 billion copyright settlement in which AI firm Anthropic will pay authors about $3,000 per book after using pirated copies to train its Claude chatbot. District Judge Araceli Martínez‑Olguín ruled the class‑action settlement provides “meaningful relief” to affected creators. Of the more than 482,000 books covered, roughly 91 % have been claimed, triggering payments to authors and publishers. The case originated with novelist Andrea Bartz and two co‑plaintiffs in 2024. An earlier preliminary approval was issued by Judge William Alsup, who had previously held that AI training on copyrighted works is not illegal per se but that Anthropic’s acquisition of millions of books from pirate sites was wrongful. Anthropic’s deputy general counsel cited the ruling as confirming that training AI on books can fall under fair‑use doctrine. This marks the largest known copyright recovery in history and the first major settlement among ongoing AI copyright lawsuits.