hot events · 2026-05-17

vLLM exposes the ugly truth of local multi-GPU inference: heterogeneous rigs work only if the engine handles the pipeline sanely. On Qwen3.5-397B-A17B with 75k tokens across seven mixed Blackwell/Ada GPUs, vLLM hits 9.8s TTFT and 7,683 t/s. llama.cpp lands at 57.2s and 1,319 t/s, roughly a 6x gap. The useful detail is not “vLLM is faster.” It is manual layer placement via VLLM_PP_LAYER_PARTITION, which balances fast Blackwell cards against slower 4090s doing FP4 emulation. SGLang looks fine on pure Blackwell, with 5.3s versus vLLM’s 5.0s on Qwen3.5-122B, then crashes when Ada enters because FP4 lacks a software fallback. Single Reddit benchmark, single topology, no independent replication; still, anyone stitching together used 4090s for 397B-class models should take the warning seriously.

A 24 tok/s LFM2-24b-a2b run puts Android local inference inside the usable zone. The reported setup is concrete: Honor Magic 7 Pro, Snapdragon 8 Elite, 24GB RAM, Q4 quantization. Gemma lands around 11 tok/s, while the MoE model reportedly hits about 24 tok/s. The wild part is CPU beating NPU and GPU in that setup. Qualcomm has sold the AI Engine story for years, but LocalLLaMA-style tests keep exposing the boring layer: memory movement, operator coverage, and runtime glue. The Reddit page is blocked by 403, so batch size, context length, backend, and sampling settings are not available here. I read this as a good sign for on-device MoE, and a bad sign for the claim that phone NPUs automatically own LLM inference.

Semble matters because it attacks the boring cost center in coding agents: context waste. On a 63-repo benchmark, it claims 98% fewer tokens than grep+read, 0.854 NDCG@10, and roughly 1.5 ms CPU queries. The stack is not magic: Model2Vec embeddings, BM25, RRF fusion, then reranking. I don’t buy grep+read as the serious opponent. Cursor, Claude Code, and Sourcegraph Cody have moved past naked grep into repo maps, AST-ish indexes, and symbol search. Still, the direction is right. Coding agents fail less from “not enough intelligence” than from retrieving 40 bad chunks and spending the next two calls laundering that noise.

MiroThinker-1.7 mini has a clean pitch: 30B total parameters, 3B active, Qwen3 MoE base, weights on HuggingFace. That is not a leaderboard flex. It is an attempt to squeeze a deep-research agent into hardware people actually own. Sliding window K=5 plus episode restarts also admits the hard part: long research runs still break context, so the system is patching continuity with control flow. Reddit is 403-blocked here, so benchmark scores, tool success rate, VRAM use, and tokens/sec are not visible. The LocalLLaMA question about consumer hardware speed is the right pressure test. DeepSeek-R1-Distill and Qwen3 already lowered the “can run locally” bar; MiroThinker needs to show a research loop that stays usable on 24GB or 48GB cards, not just another open-weight badge.

Abliterlitics hits the uncomfortable layer: refusal behavior can be stripped as an engineering target, not argued around in policy docs. The disclosed hooks are concrete enough: 85 GPU-hours on Qwen3.6-27B, five abliteration variants, HarmBench, KL divergence, and weight forensics. The summary says all five reached near-complete safety removal; Huihui kept the smallest benchmark deltas, while Heretic had the lowest KL divergence. Reddit blocked the body with a 403, so I cannot verify exact scores, sample sizes, or reproduction scripts. Still, the pattern is clear. LocalLLaMA is moving from “which prompt bypass works” to “which weight edit preserves capability while deleting refusals.” That is a much nastier problem for open-weight safety than another jailbreak leaderboard.

TGO matters because it treats visual feedback as scores, not forced winner/loser pairs. The mechanism is simple: a score-distribution threshold sets update direction, and distance from that threshold sets weight. The paper tests Stable Diffusion v1.5, FLUX, Wan 1.3B, and Meissonic, so this is broader than a one-backbone diffusion trick. I don’t buy the “new paradigm” framing. PMPO already loosens unpaired positive/negative feedback, and QRPO handles pointwise absolute rewards through quantiles. TGO is the visual-generation engineering compromise: cheap, readable, and easy to plug into diffusion or masked generators. The weak spot is the global threshold. It compresses prompt difficulty, style taste, and reward-model bias into one cutoff. If the scorer is skewed, pseudo-negatives will suppress minority aesthetics with mathematical confidence.

Suleyman’s “18 months to automate everyone sitting at a computer” is too clean for the evidence given. The snippet names accounting, law, marketing, and project management, but gives no benchmark, error rate, liability model, or deployment cost. The hard part in white-collar work is not drafting a document. It is context access, approvals, audit trails, system permissions, and owning bad calls. A Microsoft AI CEO talking up “superintelligence” is expected. Compressing the timeline to 18 months is the aggressive part. OpenAI, Anthropic, and Google are already pushing agents into Office, IDEs, support, and analytics, but task automation and job replacement are separated by procurement, compliance, and accountability. I don’t buy this claim without reproducible enterprise agent success rates.

Grok Imagine is leading with X distribution, not model evidence. The post says it is available to all users, supports realistic text-to-image output, and offers multiple aspect ratios. It gives no pricing, regional limits, model card, safety policy, or reproducible comparison against Midjourney, GPT-4o image, or Imagen. That omission matters because image generation is already crowded and heavily benchmark-resistant. The wild part is the channel. X gives Grok a default creation-and-sharing loop that standalone image tools have to buy through ads or creator communities. Even a second-tier model can absorb casual meme, avatar, and post-illustration demand if the button sits inside the feed. I don’t buy the implied capability claim until we see hard prompts: text rendering, character consistency, editing control, and commercial-use terms. Right now the product surface is visible; the moat is not.

DeepSeek V4’s 1M context is not proving whole-repo coding here; it shows a usable band. The user tested 45k, 180k, and 520k-token codebases. Their sweet spot was 150k-250k tokens. Past 300k, line-number precision degraded. At 520k, the model shifted into architecture summaries and skipped implementation details. I trust that Reddit failure mode more than the 1M headline. Coding needs retrieval, references, and local edits, not a giant prompt stuffed with a repo. Gemini 1.5 Pro had the same 1M-context aura, and serious users still leaned on chunking, search, and repo maps for reliability. The body is blocked by 403, so prompt, model settings, and DeepInfra config are missing. But the “long enough becomes a summarizer” pattern is painfully familiar.

FT’s claim reads like a product judgment, not a model judgment. The disclosed text names ByteDance and Kuaishou, limits the claim to advertising and entertainment video, and gives no benchmark, model version, blind-test size, or target against Sora, Veo, or Runway. That is too thin for “pull ahead of US rivals.” I still buy half the direction. Video generation is not won by a single demo clip; it rewards asset pools, creator feedback, ad conversion loops, and moderation pipes. Douyin and Kuaishou have daily short-video production and ad testing loops that pure model labs do not. OpenAI Sora owns the launch-stage perception; ByteDance and Kuaishou are closer to commercial quality tuned through gray-release A/B tests. Until metrics show up, read this as platform production advantage, not proof that Chinese video models beat US models.

Agentic coding’s ugly limit is not patch generation; it is hidden reasoning spend. The visible summary gives a sharp hook: OpenHands traced 8 frontier models on 500 SWE-bench-verified tasks, with a 154:1 input-output token ratio and only 0.32 Kendall tau between human difficulty labels and token use. Human “hard” does not predict token burn cleanly, which is exactly the failure mode vendors avoid showing in demos. That hits the margin story behind Cursor, Devin, and OpenHands-style workflows. A higher SWE-bench pass rate looks great on a launch slide; enterprise buyers care about cost per merged PR. The full WeChat body is blocked by verification, so model names and pricing assumptions are not disclosed. I’d treat 154:1 as a serious warning flare, not a settled measurement.

The 603B-token number matters less than OpenAI eating a $1.3M bill for Peter Steinberger. Over 30 days, 7.6M requests averages about 79K tokens per request, which smells like continuous code-agent traffic, not chat usage. His claim that disabling fast mode cuts price by 70% says latency is becoming a hidden tax in agent products. I don’t read this as generosity. It looks like OpenAI buying a real workload trace from a very visible power user. The article body is only a CAPTCHA page, so model name, cache hit rate, and input/output split are not disclosed. Without those three, the $1.3M figure proves burn rate, not viable unit economics.

The world-model narrative is getting financially front-run. More than $10B over 18 months sounds like consensus, but it also bundles robotics, video generation, simulation, and agent training into one investable label. The hard hook is DreamDojo: 44,711 hours of first-person video pretraining, then r=0.995 correlation between policy evaluation and real robot outcomes. If that holds, it moves from “predicts frames” toward “filters robot policies.” I don’t buy the number at face value yet. The available body is a CAPTCHA page, so the benchmark setup, task mix, robot hardware, and correlation method are not disclosed. r=0.995 is unusually clean for robotics. NVIDIA Cosmos, Genie-style environment models, and LeCun’s JEPA line are all circling this terrain; the useful test is transfer across embodiments and long-horizon failure, not whether the deck says “world model.”

Dario’s framing is careful: admit high unemployment and inequality first, then put 5%-10% GDP growth on the table. That gives Anthropic room to sell automation without pretending it is just a safety lab watching from the sidelines. The hard claim is near-free software cost; that hits SaaS seats, implementation fees, and outsourced dev work in one shot. I don’t buy the clean “engineers move into editing or upgrading work” line. Claude Code, Cursor, and Devin already show the editor layer does not appear one-for-one for displaced engineers. AI compresses delivery from billable human time into task output; bargaining power falls before neat new roles arrive.

Amodei is overstating the “software becomes near-free” line. The body gives only a WSJ YouTube interview and the claim that million-user cost spreading breaks; it gives no price curve, timeline, or software category. SaaS cost is not code alone. Compliance, sales, integration, uptime, and liability remain stubbornly non-free. Anthropic still prices Claude by tokens, and enterprise AI still sells permissions, audit logs, and data isolation. Code generation will crush prices for CRUD apps, internal tools, and disposable scripts; that part is real. But Workday, ServiceNow, and Salesforce customers buy workflow ownership and risk transfer. Amodei’s warning works as a labor-market alarm. It fails as a clean forecast for software margins going to zero.

Vibe coding’s security failure sits in deployment defaults, not model-generated code. The snippet says “thousands” of enterprise apps were exposed, including hospital schedules, bank financial data, and clinical trial data. That is not a toy-app bug; that is regulated data on the open internet. I don’t buy the easy excuse that users misconfigured access. Tools like Lovable, Replit Agent, and Bolt push non-engineers straight from prompt to production, so the default permission becomes the security boundary. The body does not disclose the named platforms, exposure duration, or remediation path, and those gaps matter. But the mechanism is already damning: AI code review will not catch a public-by-default deploy button, and enterprise procurement often misses that layer.