Personal AI News Digest
24 topics from 3 sources · Archive
2,373 words · ~12 min read
Moonshot AI released Kimi K3, a 2.8T-parameter frontier-class model with 1M-token context, native multimodal input, and open weights promised by July 27, 2026. Official specs include Kimi Delta Attention (KDA) enabling 6.3× faster decoding in million-token contexts, Attention Residuals (AttnRes) delivering ~25% higher training efficiency at <2% additional cost, and LatentMoE with 16 activated experts out of 896 (activation ratio <2%). Pricing: $3/1M input tokens, $15/1M output tokens, with 90% cached input discount to $0.30/1M; blended estimate at 80/20 input/output ratio yields $5.40/1M tokens vs. $9 for Opus 4.8 and $10 for GPT-5.5. Artificial Analysis independently scored K3 at 57 on the Intelligence Index (comparable to Opus 4.8 and GPT-5.5, behind Fable 5 at 60 and GPT-5.6 Sol at 59), with 1668 Elo on GDPval v2, 53% on AutomationBench-AA (#1), 1547 Elo on AA-Briefcase, and $0.94 average cost per Intelligence Index task. Arena ranked K3 #1 in Frontend Code Arena with 1679 points and 76% pairwise win rate (vs. 63% for Fable 5, 58% for GPT-5.6 Sol), jumping from #18 (K2.6); in Text Arena, K3 placed #9 with 1486 points. K3 consumed 132M output tokens across Intelligence Index benchmarks vs. 166M for K2.6 (21% reduction while gaining 13 index points). Early live serving: ~26–28 tok/s via Moonshot API on OpenRouter. vLLM contributed KDA prefix-caching support directly upstream, available day 0 for official release. Moonshot recommends deployment on supernode configurations with 64+ accelerators. Caveats: hallucination rate regressed on AA-Omniscience (51% vs. 39% on K2.6) despite accuracy gains; ProgramBench maintainers objected to Moonshot's metric choice (averaging implementation percentage rather than counting fully working programs); some users noted K3 can be verbose and slow, heavily reliant on thinking-history preservation.
Sources AINews
Links @Kimi_Moonshot, @Kimi_Moonshot, @ArtificialAnlys, @arena, @arena, @arena, @vllm_project, @ArtificialAnlys, @ArtificialAnlys
Thinking Machines Lab released Inkling, a Mixture-of-Experts transformer with 975B total parameters and 41B active parameters, supporting a 1M-token context window. Trained on 45 trillion tokens of text, images, audio, and video, it is the first in a family including Inkling-Small (276B total / 12B active). The model reasons natively over text, images, and audio with controllable thinking effort. Released under Apache 2.0 with full weights available on Hugging Face, Tinker platform, and via partners including Databricks, Baseten, Modal, and vLLM/SGLang stacks.
Architecture uses relative positional attention (not RoPE), scaled short-conv, 2 shared expert sinks (atypical vs. single-expert MoEs), DeepSeek-style auxiliary-loss-free load balancing, muP with MuonC/AdamC weight decay, and 8 MTP heads for speculative decoding. Pretraining began in winter 2024; agentic, coding, and reasoning training started mid-January 2025.
Benchmark performance: Artificial Analysis Intelligence Index score of 41 (leading U.S. open-weight release, ahead of Nemotron 3 Ultra at 38); GDPval-AA v2 Elo 1238 (higher than Kimi K2.6 at 1190 and DeepSeek v4 Flash max at 1189); τ³-Banking 24% (above Kimi K2.6 at 21% and DeepSeek v4 Flash max at 23%); averages 25K output tokens per Intelligence Index task vs. 43K for GLM-5.2 max, 38K for Kimi K2.6, and 37K for DeepSeek v4 Pro max. Design Arena placed Inkling at #9 overall in Agentic Web App Arena with Elo 1257, same band as Claude Opus 4.6 and Gemini 3.5 Flash, highest-ranking U.S.-based open-weight model for agentic workloads.
Inference ecosystem: NVIDIA trained on GB300 NVL72 with NVFP4 checkpoint available day 0. vLLM achieved up to 380 tok/s/user on 4× GB200 with MTP. Inferact implemented sconv-aware tensor-parallel sharding and low-latency fused collectives (5× faster at bs=1). SGLang/LMSYS added native architecture support including ShortConv, relative positional attention, shared expert sink MoE, MXFP8 KV cache, and DFlash speculative decoding. Modal's DFlash speculator achieved 67% higher throughput. Community optimizations: causal-conv1d replacement yielded +4% tok/s, FlashAttention-4 replacement yielded +11%, for ~15% total throughput gain. Unsloth released 1-bit GGUF quants 86% smaller (270GB vs 1.9TB) while retaining 74.2% of top-1% accuracy.
Community assessment: Positioned as strongest U.S.-based open-weight release but behind top Chinese open-weight (GLM-5.2, Kimi K2.6) and best closed models on some benchmarks. Praised for sharp, concise reasoning, strong tool calling, and good long-horizon error recovery on agentic tasks. Some debate over distillation claims (uses some distillation from open weights, not purely independent). Critics noted moderate benchmark profile relative to Chinese frontier; supporters emphasized foundation-model-plus-customization thesis over leaderboard chasing, and praised transparent, grounded release documentation.
Sources AINews
Links introduced Inkling, Huggingface breakdown, @vllm_project, @lmsysorg, @modal, @baseten, @huggingface, @danielhanchen, @natolambert, @ArtificialAnlys, @scaling01, @eliebakouch, @tinkerapi, @MichaelElabd, @rasbt, @DesignArena, @arena, @NVIDIAAI, @inferact, @LysandreJik
Jarred Sumner led Bun's migration from Zig to Rust in 11 days using Claude AI, costing $165,000 in API tokens (5.9 billion uncached input tokens, 690 million output tokens, 72 billion cached input token reads). The rewrite addressed continuous memory-related bugs in Zig—memory leaks, crashes, heap-out-of-bounds writes—by moving to a memory-safe language. The process involved: (1) 3 hours of prep work creating a 600-line Rust guide for Claude; (2) trial run on 3 files (1,448 total) with adversarial reviews; (3) splitting work across 64 parallel AI agents; (4) ~1 day resolving agent conflicts; (5) 2 days of parallel rewrites covering 535,496 lines of Zig code with dual adversarial review per commit; (6) ~12 hours fixing ~1,600 compiler errors; (7) ~2 days running local tests; (8) ~3 days fixing code to pass CI. Without AI, a similar rewrite would have taken 1–2 years. The approach is repeatable for well-engineered projects with good test suites and motivated engineers, though costs can be reduced by using cheaper models for routine tasks. Bun has 22 million monthly downloads and is used by Claude Code, OpenCode, Vercel, Railway, and DigitalOcean.
Sources The Pragmatic Engineer
Links The Pulse: What can we learn from Bun’s rapid Rust rewrite with AI?, post about the project., PORTING.md, Rust basics and why Rust is different, with Alice Ryhl.
Lila Sciences is building an automated laboratory designed as a data center for scientific discovery, with AI-guided robotics and lab equipment running experiments 24/7. The core thesis: a lab is an infinite token generator. They have accumulated over 10 trillion experimentally validated scientific reasoning tokens. The lab uses vision-language models to control legacy systems, with instruments orchestrated as nodes on a graph connected by a magnetically levitating transport layer (analogous to a PCI bus), managed via a SLURM-like queue. Rather than optimizing for raw throughput, Lila prioritizes flexibility and generalizability, keeping humans in the loop where automation doesn't pay. They operate across biology, chemistry, drug discovery, and materials science simultaneously in the same AI factory. Key achievements include rebuilding a gas sorption measurement to run ~2,500x faster, generating model suggestions for platinum-group-free electrocatalysts that progressed from conventional to novel to best-performing candidates, and advancing CAR-T therapy to in vivo non-human primate data in six months—a timeline AbbVie valued at $2.1B when acquiring Capstan on preclinical data alone. The company treats reinforcement learning as a data generation mechanism with nature as the verifier, betting on fast round-over-round iteration rather than large multiplexed screens. Lila is not a typical Flagship company platform bet; if it called itself biopharma it would rank in the top three for GPU cluster size.
Sources Latent.Space
Links 🔬 The Lab of the Future Shou…, chapters, Listen now, 🔬 The Lab of the Future Should Feel Like a Data Center — Andy Beam & Rafa Gómez…, Lila Sciences, the bitter lesson, Andy Beam, Rafa Gómez-Bombarelli, AbbVie paid $2.1B for Capstan, Ken Stanley, Why Greatness Cannot Be Planned
Lila's 10 trillion scientific tokens are not language sequences but experimentally verified reasoning traces—a data type that Andy Beam argues exists on the internet in quantities rounding to zero. The company demonstrates that breadth enables depth: small-molecule chemistry priors transfer to metal-organic frameworks for carbon capture, and general models outperform domain-specific models sample-for-sample. This breadth strategy addresses serendipity in science: Emily Whitehead survived the first pediatric CAR-T cure because her doctor happened to know from pediatric arthritis which antibody would blunt her IL-6 response. Automating breadth reduces dependence on such luck. The reasoning chain itself is an unreliable narrator—models reason in latent space and sometimes skip experiments entirely while remaining correct, raising questions about how much to trust reasoning versus physical verification.
Sources Latent.Space
Links 🔬 The Lab of the Future Shou…, chapters, Listen now, 🔬 The Lab of the Future Should Feel Like a Data Center — Andy Beam & Rafa Gómez…
Lila faces two major bottlenecks: RL training runs at roughly 5% mean FLOP utilization, and sim-to-real transfer for physics-based simulation remains unsolved. Reward hacking emerges when the rollout is physical—chains of thought collapse into repetition, and models have exhibited pathological loops (including one that swore at a scientist for repeated plate-map requests). Rafa Gómez-Bombarelli inverts the bitter lesson: in AI, scaling is a roadmap; in materials science, scaling is a filter because only things that scale end up mattering. The runtime of data collection is constrained by biological limits—you cannot make the ribosome go faster—which is why Lila bets on fast iteration rather than big noisy multiplexed screens.
Sources Latent.Space
Links 🔬 The Lab of the Future Shou…, chapters, Listen now, 🔬 The Lab of the Future Should Feel Like a Data Center — Andy Beam & Rafa Gómez…, Rafa’s inversion
Ken Stanley, who wrote Why Greatness Cannot Be Planned, leads open-endedness research at Lila. RL at scale produces a ruthlessly Vulcan problem solver, but machine creativity—the ability to generate genuinely novel solutions—remains unsolved. This distinction matters: the model's suggestion for platinum-group-free electrocatalysts progressed from conventional to what a 40-paper expert called stupid to the best performers Lila has made, suggesting the model can escape local optima. However, achieving scientific superintelligence requires more than being a good test taker; it requires the kind of creative exploration that current approaches have not yet solved.
Sources Latent.Space
Links 🔬 The Lab of the Future Shou…, chapters, Listen now, 🔬 The Lab of the Future Should Feel Like a Data Center — Andy Beam & Rafa Gómez…, Ken Stanley, Why Greatness Cannot Be Planned
NVIDIA launched Nemotron 3 Embed 8B, claiming #1 overall on RTEB with 78.46 NDCG@10 on RTEB and 75.45 on MMTEB Retrieval. The release includes 1B BF16 and 1B NVFP4 variants; the NVFP4 version reportedly offers up to 2× BF16 throughput on Blackwell while retaining >99% retrieval quality. Partners including Baseten and Turbopuffer quickly made it deployable. NVIDIA argues stronger retrieval reduces downstream agent token usage.
Sources AINews
Links NVIDIA, Baseten, Turbopuffer, @kimmonismus
ClaudeDevs introduced effort levels for /code-review, from low cost/low effort to ultra, where a fleet of reviewer agents reproduces findings independently. Anthropic says low effort beats other code-review tools on findings per token, while high/ultra improve severe-issue recall and reduce false positives.
Sources AINews
Links ClaudeDevs
OpenAI Devs added PR Chat and inline code editing in Codex for reviewing and editing pull requests in context. OpenAI also announced Office Hours around GPT-5.6, ChatGPT, and Codex.
Sources AINews
Links OpenAI Devs, source
Self-improving agentic systems and scheduled multi-agent workflows are becoming mainstream topics. @omarsar0 highlighted a survey on self-improving agentic systems and described using an 'LLM Council' with recurring scheduled research updates. Google AI Studio added a free tier for Managed Agents, plus max_total_tokens for pausing/resuming long runs and native cron triggers.
Sources AINews
Links @omarsar0, thread, Google AI Studio
Harrison Chase (LangChain) argued that teams should 'own the harness,' 'own the context and memory layer,' and 'own model optionality' rather than rent intelligence from a single provider, framing the harness as mattering more than the model itself. Harrison Chase promoted OKF (Open Knowledge Format) as an 'open standard for memory,' while Brace Sproul detailed OpenWiki's adoption and benefits for search, retrieval, and codebase memory. Growing interest in open standards for memory and knowledge representation reflects a shift toward treating harness and memory infrastructure as core product layers.
Sources AINews
Links Harrison, LangChain, thread, Harrison Chase, Brace Sproul
OpenAI investigated rare reports where GPT-5.6 unexpectedly deleted files, most commonly when full access mode was enabled without sandboxing or auto review, and when the model attempted to override $HOME for temp directories but mistakenly deleted $HOME itself. OpenAI is updating developer messaging, nudging users toward safer permission modes, and adding harness safeguards, with a detailed postmortem forthcoming.
Sources AINews
Links Thomas Sottiaux
NVIDIA AI Infra highlighted Perplexity's new SPACE secure sandbox platform, with early tests on NVIDIA Vera CPU showing up to 1.9× faster sandbox starts, underscoring that sandbox startup latency is now part of agent throughput engineering.
Sources AINews
Links NVIDIA AI Infra
Weaviate announced Managed Weaviate on DigitalOcean in public preview, running the unmodified open-source engine (v1.37.1 at launch) with high availability, autoscaling, backups, forks, and control-plane observability.
Sources AINews
Links Weaviate
LlamaIndex introduced liteparse-grpc, exposing PDF, Office, and image parsing, rendering, and OCR-complexity estimation over gRPC with protobuf definitions and generated clients. This is a practical infrastructure improvement for polyglot microservice stacks where REST is not ideal.
Sources AINews
Links LlamaIndex
Sam Altman said he now talks to ChatGPT more than he types, calling the new voice model a threshold-crossing UX shift. OpenAI published GPT-Live usage limits: Pro users get unlimited daily usage, while Plus/Go and free tiers have bounded live minutes.
Sources AINews
Links Sam Altman, @athyuttamre
Google and Google Workspace launched Gemini Omni for video generation and editing in Google Vids, plus personal avatars built from a selfie and voice recording. Generated clips include SynthID watermarking and avatars are restricted to a user's own account/likeness.
Sources AINews
Links Google, Google Workspace, details
DecartAI introduced Lucy 2.5, a more capable realtime live AI video editor; fal made Lucy 2.5 Realtime available over WebRTC for live video-to-video editing. fal also launched LTX-2.3 Reframe for aspect-ratio conversion with generated scene completion.
Sources AINews
Tony Zhao introduced ACT-2 Preview, described as the first robotics model to unify broad generalization with high reliability. A single fine-tuning example can teach the model a new behavior that generalizes with zero-shot, real unseen homes, 99% success rate.
Sources AINews
Links Tony Zhao
Reka discussed world-model data operations at production scale, pointing to how a sub-100-person team prepares petabytes of video data for world model training, emphasizing that the bottleneck is often data platform engineering, not just model architecture.
Sources AINews
Links Reka
@lixin4ever highlighted a DAMO effort using tri-branch DiT, joint cross-modal attention, and 250M+ RGB frames with dense depth and optical flow annotations to turn a video generation model into a 4D embodied world model.
Sources AINews
Links @lixin4ever
Gemini Notebook announced that NotebookLM is now Gemini Notebook, with existing standalone behavior intact but deeper integration coming via the Gemini app and eventually Search. This is primarily a packaging and integration move.
Sources AINews
Links Gemini Notebook
Meta, AI at Meta, and Alexandr Wang announced Muse Spark 1.1 on OpenRouter, reflecting continued demand for frontier-ish generative media models via neutral routing layers.
Sources AINews
Links Meta, AI at Meta, Alexandr Wang