HackerNews Digest

Open‑source large language models have reached near‑commercial quality, but their parameter counts now require massive GPU memory. Kimi K3 (2.8 T parameters, >1.5 TB VRAM for weights + 1 M‑token KV cache) cannot fit on an 8‑GPU NVIDIA B200 node; it must run on larger B300 nodes (288 GB GPU⁻¹) or on AMD MI355X GPUs, which offer comparable HBM capacity at ~2.4× lower price per GPU. Benchmarks on a 1 024‑token input/400‑token output task show the MI355X achieving 952 tokens /s /node (118 tokens /s single‑stream), 3.8× the aggregate throughput of a two‑node B200 (TP16) deployment and 1.3× the single‑stream decode rate, while B300 nodes remain 1.65× faster in aggregate but cost 2.4× more. Performance‑per‑dollar is $2.50 / GPU‑hr (MI355X) vs $4.25 (B200) and $6.00 (B300).

Key engineering improvements: * Fixed missing top_k_renorm_prob function in ROCm’s sglang sampling path, yielding ~2.2× single‑stream and ~1.7× moderate‑load decode speedup and 18 % higher peak aggregate throughput. * Resolved a head‑count shape mismatch for AITER MLA prefill kernel (zero‑pad 12→16 heads), enabling fast ASM kernel use and accelerating 172 k‑token cold prefill from ~51 s to ~23 s (2–3× faster) without affecting decode speed.

These results suggest AMD GPUs can match or exceed NVIDIA in cost‑efficient serving of frontier open‑source models.

Go 1.27 introduces several language and runtime enhancements:

• Generic methods: methods can declare their own type parameters independent of the receiver, enabling constructs like func (b Box[T]) Map[U any](f func(T) U) Box[U]. Interfaces still cannot contain generic methods.
• Struct literal selectors: promoted fields from embedded structs can be set directly in literals (e.g., User{ID:7, Name:"Mittens}).
• Generalized function type inference: generic functions can be used in any context expecting a matching function type (assignments, conversions, composite literals) without explicit type arguments.
• Size‑specialized allocation: the compiler emits specialized malloc calls for allocations ≤ 80 bytes, reducing cost up to 30 % (≈ 1 % overall), with a 60 KB binary size increase. Opt‑out via GOEXPERIMENT=nosizespecializedmalloc.
• Goroutine labels in tracebacks: runtime/pprof labels now appear in panic, SIGQUIT, and runtime.Stack outputs; disable with GODEBUG=tracebacklabels=0.
• Goroutine leak profile: a built‑in runtime/pprof “goroutineleak” profile detects permanently blocked goroutines without experimental flags.
• Post‑quantum signatures: new crypto/mldsa implements ML‑DSA (parameter sets 44, 65, 87) and integrates with crypto/x509 and TLS 1.3.
• Standard UUID package: uuid supports RFC 9562 parsing, generation (New, NewV4, NewV7), and comparison.
• JSON v2 default: encoding/json now runs on the v2 implementation; behavior unchanged, with optional GOEXPERIMENT=nojsonv2 to revert.
• Portable SIMD: experimental simd package offers vector‑size‑agnostic SIMD types (Int32s, Float32s, etc.) activated via GOEXPERIMENT=simd.
• CutLast functions: strings.CutLast and bytes.CutLast split on the last occurrence of a separator.
• Generic hashing: hash/maphash adds Hasher[T] interface (e.g., case‑insensitive string hasher) for custom hash/equality strategies.

Seedance 2.5, the latest video‑creation model from ByteDance’s Seed team, expands single‑pass generation to 30 seconds and supports multi‑round extensions, enabling multi‑minute stories with consistent audiovisual language and smoother shot transitions. The model accepts up to 30 images, 10 video clips and 10 audio clips per request, improving reference‑based generation for complex scenes, multi‑character interactions, and specific styles such as clay‑render or motion cues. Editing is refined through timestamp‑level control, allowing precise adjustments of actions, camera moves, green‑screen replacements, and perspective changes while preserving continuity and realistic lighting, texture, and physics. Seedance 2.5 is positioned for professional use in film, advertising, education, manufacturing, and autonomous‑driving simulations, where it can produce synthetic training data and immersive instructional visuals. The service is currently available on Jimeng AI, Doubao Pro, with API access planned via BytePlus ModelArk. Ongoing work targets improved physical plausibility, more intuitive generation, and broader industry integration.

The page is titled “Diátaxis.” It presents a navigation element labeled “Next” and a link or prompt titled “Start here – Diátaxis in five minutes,” suggesting an introductory resource. Under a section titled “Images and Visual Content,” two images are referenced: the first with alt text “Logo,” and the second with alt text “Diátaxis.” No additional textual description, explanatory content, or technical details are provided beyond these headings and image placeholders. The layout consists solely of the title, navigation cue, introductory link, and the two image descriptors.

The guide details configuring an ASRock BC‑250 “budget steam machine.” It recommends allocating 6 GB of UMA VRAM via BIOS (UMA_MODE = UMA_SPECIFIED, UMA_FRAME_BUFFER_SIZE = 6 GB) rather than the common 512 MB setting. For GPU power and performance, install the cyan‑skillfish‑smu governor (via AUR on CachyOS or rpm‑ostree on Bazzite), adjust the config /etc/cyan‑skillfish‑smu/config.toml to set a frequency range of 350–1850 MHz (or up to 2000 MHz if extra compute units are not enabled), lower the minimum frequency, and tweak load‑target percentages and safe‑point voltages in 50 mV steps, testing stability after each change. Enable the 40‑CU unlock with the bc250‑cu‑live‑manager script, activating all functional compute units and installing the service to apply on boot. For CPU overclock/undervolt, install bc250‑smu‑oc, provide the “stress” binary, then run bc250‑detect with target frequencies (e.g., 4000 MHz) and VID values (≈‑34 mV) to generate an overclock.conf. Finally, increase the shader‑cache limit by adding MESSA_SHADER_CACHE_MAX_SIZE=5G to /etc/environment and disable Steam’s pre‑cached shader download to avoid repeated redownloads.

MIT Sloan researchers evaluated financial advice from large language models (LLMs) such as GPT‑5.x and Gemini 3 Flash. Using a life‑cycle simulation of income, taxes, and investments, they asked 1,000 adults to submit prompts for spending and investing advice, then simulated outcomes for ages 22‑89 if the advice were followed. Results showed LLMs consistently recommended higher savings, diversified stock exposure, and age‑appropriate risk reductions, producing sizable wealth buffers for most users above age 30. Structured, “academic” prompts that included detailed personal and economic assumptions yielded better outcomes than typical user prompts. However, the models relied on simple heuristics, failed to rebalance portfolios actively, and responded differently based on the prompter’s gender, financial literacy, and AI experience, creating wealth gaps of 4‑6 % at retirement. The study suggests AI can offer low‑cost, broadly accessible guidance but requires careful prompting and awareness of bias, and may complement rather than replace human financial advisors.

Kyoto University researchers developed a rapid‑training protocol that induces habit formation in mice within days, enabling longitudinal assessment of behavioral and neural changes. Using a two‑stage regimen—initial goal‑directed training followed by four days of habit‑facilitating training—they identified two distinct cortico‑striatal circuits that regulate habit acquisition and execution.

• The anterior cingulate cortex (ACC) → retrosplenial cortex pathway weakens during the transition, determining whether a behavior becomes habitual.
• The lateral orbitofrontal cortex (lOFC) → central striatum pathway modulates the magnitude of habit execution; stronger neural responses correspond to higher behavioral output, while weaker responses reduce execution.

Optogenetic manipulation of these pathways selectively promoted habit formation or altered execution levels, confirming their separate functions. The findings reveal that habit formation is not a simple replication of repeated actions but involves circuit‑specific modulation, accounting for individual differences in habit intensity. Future work will explore mechanisms underlying these differences, with potential implications for enhancing beneficial habits and treating compulsive‑behavior disorders.

RFC 10015 deprecates obsolete key‑exchange methods in TLS 1.2 and DTLS 1.2. It identifies security and interoperability issues with finite‑field DH (FFDHE/FFD H) and RSA key exchange:

• Finite‑field DH (FFDHE): lacks group‑negotiation, leading to incompatibility; custom or small groups may contain weak sub‑groups; many implementations still use 1024‑bit groups, providing only a narrow margin against the current 795‑bit discrete‑log record; a few large‑scale computations can break traffic using standardized groups; non‑ephemeral or partially‑ephemeral use is vulnerable to the Raccoon timing side‑channel unless constant‑time mitigations are applied.

• RSA key exchange: provides no forward secrecy; susceptible to Bleichenbacher‑type attacks (e.g., ROBOT, DROWN) because correct countermeasures are hard to implement; lacks domain separation, so a compromised RSA key endangers all endpoints sharing it.

The document updates multiple TLS‑related RFCs, aligns with BCP 195, and specifies deprecation of the affected cipher suites in sections 5.2‑5.5. All other BCP recommendations remain unchanged.