Highlights
What happened at the companies we cover, by day, newest first. Every item links to its sources.
Showing 1–20 of 42 highlights
Friday, Nov 13 42 stories
Three major outages hit Kimi K3.1 Pro on Together AI in one day
ComplaintXBluesky
Together AI's status feed reported three separate major outages for its Kimi K3.1 Pro endpoint on Nov 13. The incidents were marked resolved after roughly 36 minutes, 3 minutes and 21 minutes respectively, with no resolution time given while the third was open. Monitoring accounts relayed each start and resolution on X and Bluesky, though there was little direct user commentary. Three failures in one day on a single popular open-weight model is a reliability signal for teams depending on that endpoint.
Sources
2 more sources
Gamma details how splitting its agent into jobs let small models cut deck cost to cents
OtherXOther
Gamma co-founder Jon Noronha published the first Gamma Research post in November, explaining how the Gamma 5 agent moved from one frontier model, which cost $1.30 to $7.40 per deck in early prototypes, to many small models handling separate jobs such as planning, template choice, writing, repair and review. Slides are rendered in a browser and checked by hundreds of programmatic rules, with a model called back only when a check fails, and each job is A/B tested on live traffic, with more than twenty model experiments in October. Gamma found no single best small model, uses Haiku for edits and GPT-6 Luna for slide repair, and says a small repair model it trained fixed 76% of failing slides on the first try versus 57% for a general model. The post spread on X with positive reactions focused on the cost drop.
Sources
Users keep comparing Instinct to Muse and Dot, with a data-terms complaint
DiscussionX
On Nov 13, X users kept ranking Instinct against other personal agents such as Meta's Muse, ChatGPT Dot and Claude Duet, asking which platform works best. One user said Muse was the only usable option and called Instinct's data terms shady, while others praised how much their Instinct agent handles for them. An AI newsletter account warned businesses that agents like Muse and Instinct are already logging into sites, checking stock and trying to buy. Requests and trades for Instinct invite codes also continued, a sign that access is still gated.
Sources
2 more sources
ElevenCreative pushes Dub & Sync, one-step dubbing with lip sync across 25 languages
OtherX
ElevenLabs' ElevenCreative account announced on Nov 12 that Dub & Sync is live and spent Nov 13 promoting it. Users upload a video, choose target languages from a list of 25, and get back a dubbed version with the speaker's lips synced to the new audio, without separate editing or prompting. The company pitches it as solving the mismatch between mouth movement and translated speech that makes dubbed video look off. The launch post drew large reach on X, and the follow-up promos drew moderate engagement.
Sources
ElevenLabs' ElevenAI app bundles meeting notes, dictation and agent follow-ups for teams
LaunchNewsBlueskyX
Coverage on Nov 13 detailed ElevenAI, the experimental team app ElevenLabs unveiled at its New York summit, which is available only through an early-access signup. It records and summarizes meetings while notifying participants, lets users search across past meetings and get live help during calls, and offers dictation into any app in more than 90 languages with support for industry terms. Companies can connect ElevenAgents to trigger follow-ups such as drafting emails from notes. TechCrunch and Digital Trends framed it as a push against Granola, Wispr Flow and Fireflies and a way for enterprises to skip separate notetaker licenses, and the story was widely reshared on Bluesky with mostly neutral commentary.
Sources
LangChain confirms $300M Series C at $3B and repositions around LangSmith
NewsXNewsBlueskyHN
LangChain officially announced a $300M Series C at a $3B valuation on Nov 13, led by CapitalG with GIC, DFJ Growth, IVP and Sapphire Ventures. TechCrunch notes that is more than double the $1.25B valuation from roughly a year ago. The company is centering its pitch on LangSmith as a framework-agnostic platform for running, improving and governing agents, claims more than 7,000 organizations use it, and promises a batch of new LangSmith capabilities on Nov 20. The CEO's announcement post drew over a thousand likes and reposts on X. Some replies argued the money is going to LangSmith rather than the open-source framework.
Sources
1 more source
Nscale warns it must raise more money as debt jumps to $5.5B before its IPO
NewsNewsBlueskyXHN
Nscale's amended IPO prospectus, filed Nov 12, says it will need additional financing to meet forecast 2027 obligations and that some new capital commitments cannot be deferred or cancelled. Debt rose from about $137 million at the end of June to $5.5 billion at the end of September, and Nvidia committed another $1 billion in financing beyond what was previously disclosed. A person close to the company told the WSJ it still expects to list before year-end, possibly funding needs through customer prepayments and vendor financing. Axios framed the filing as a test of whether Nvidia keeps financing customers like Nscale after its leadership change.
Sources
Hugging Face's llama.cpp team ships v1.0 of llama.cpp and ggml with stable APIs
LaunchHNXBluesky
The llama.cpp team at Hugging Face released v1.0.0 of both llama.cpp and the ggml tensor library, about three and a half years after the project began. Going forward, breaking changes to the public C API or ABI will trigger major-version bumps under explicit semantic-versioning rules, aimed at the many downstream apps that embed the libraries. The team cited support for more than 150 model architecture variants, wider packaging across Homebrew, Debian, conda-forge, Winget and Docker, and CI scaled on HF Buckets and Jobs with self-hosted hardware runners. Reaction on Hacker News, across three threads, and on X was upbeat, with commenters asking about KV-cache quantization guidance for coding and a Linux desktop app.
Sources
Sierra ships Poppy 1.0; Shopify turns it on for every store and Meta's Muse checks for it
LaunchNewsXBluesky
Sierra released version 1.0 of its Personal Agent Protocol, Poppy, on Nov 13, a standard for how personal AI agents identify themselves to companies and learn what they may do, built largely on MCP, OpenAPI and UCP with OAuth-scoped permissions. Shopify now publishes a Poppy file for every merchant automatically and co-wrote a commerce extension, Meta's Muse agent looks for Poppy first on each site, and Mercury and Turkish Airlines are live, with Synchrony working on it. Sierra says 60 new design partners joined, bringing the total past 100, including Salesforce, Intuit, Delta and Zoom, and it offers an implementation kit with a validator. The Next Web notes the spec is still labelled a draft and could change in breaking ways. X reaction was broad and mostly positive, with partner companies touting live deployments.
Sources
Vercel Sandbox boots up to 4x faster outside iad1 via regional image cache
LaunchBlueskyOther
Vercel Sandbox now reads reused image data from a regional cache, so sandboxes outside the iad1 region skip round trips to origin storage on cache hits. Caching covers managed and custom images in Vercel Container Registry, including snapshots built on them. Vercel's internal benchmarks showed startup time dropping about 75 percent in Paris and from roughly 3.5 seconds to 1.05 seconds in Seoul. Regions near origin storage see smaller gains, and the change needs no configuration; it was relayed mainly by changelog trackers.
Sources
Vercel ships Workflow SDK 5 with per-run encryption, regional runs and big latency cuts
LaunchXBlueskyNews
Vercel released Workflow SDK 5, the durable-execution library behind its Workflows product, rebuilt for long agent loops after more than 4 billion steps and 700 million runs. Every workflow and step payload is now encrypted end to end with AES-256-GCM using a per-run key, and runs can live in any Vercel Function region with failover instead of only iad1. Vercel's own benchmarks against v4 show step-to-step p75 overhead falling from 618 ms to 109 ms and a 1,020-step run dropping from 21 to 3 minutes, which Vercel stresses reflects one workload in one region. Reaction on X was broadly positive, led by Vercel's CEO, while one developer cautioned that the new run attributes are stored in plaintext.
Sources
1 more source
Runway ports Praxis-1 to low-cost SO-101 arms with an NVFP4 build for one RTX 5090
LaunchXBlueskyOther
Runway extended its open-weight Praxis-1 robot world action model on Nov 13 with checkpoints for the SO-100/SO-101 hobbyist arm, built on community datasets from the LeRobot ecosystem. The repository now lists SO-101 weights plus quantization-aware NVFP4 versions, and Runway says the SO-101 variant runs on a single RTX 5090. Praxis-1 is a 27B mixture-of-experts video-action model that predicts short video chunks and matching arm commands, and it already has a LeRobot adapter. Per the LeRobot docs, the full single-arm DROID checkpoint needs an 80 GB GPU in bfloat16, so the quantized SO-101 build lowers the hardware bar considerably.
Sources
Avianca rolls Harvey out across its multilingual legal organization
NewsOther
On Nov 13, Harvey said the airline Avianca is deploying its platform across the airline's legal organization. The team works in Spanish, Portuguese and English on aviation, regulatory and commercial matters in Colombia, El Salvador and Brazil. Lawyers will use Vault to search their own lease and maintenance clause libraries, filings and contracts, and the Word add-in to compare clauses against company standards. According to Harvey, the pilot included training for 14 legal units plus security and AI governance reviews.
Sources
Users on X call Suno v6 a downgrade from earlier models
ComplaintX
A cluster of X posts criticized Suno v6 output quality, with the most-engaged one calling the model a creative step backward with weaker melodies, generic arrangements and less control. Another user claimed, without evidence, that v6 was trained on a narrower, sanitized dataset to satisfy label partners, while others simply called it bad. The complaints surfaced alongside chatter that ACE Music v1 had passed v6 on Artificial Analysis's vocal-music leaderboard, though some users disputed that rival models beat it. Plenty of creators were still publishing songs made with v6 the same day, so sentiment was split rather than uniformly negative.
Sources
1 more source
Harvey engineering details how it enforces law-firm ethical walls across AI work
OtherX
Harvey's engineering team published a deep dive on Nov 13 on carrying law firms' ethical walls (information barriers) into AI-generated work, not just source documents. Harvey imports client matters and wall policies from firm systems through provider adapters into one common model, with matter sync and wall sync run as separate jobs. Vault files, conversations and review tables are linked to client matters, and each request checks both resource permissions and matter clearance. The company promoted the post on X, where it drew modest engagement.
Sources
Replit adds Performance Agent scans, per-Viewer sharing and Fabric CRUD apps
LaunchOther
Replit's Nov 13 changelog introduces Performance Agent, which scans an app's code and available logs for likely slowdowns from the Monitoring tab and lets users review and test fixes before publishing; it needs a paid plan and edit access but works on unpublished apps. Enterprise admins and members can now share projects with specific Viewers rather than an entire Viewer group. The Microsoft Fabric integration now supports database-backed apps that create, read, update and delete records, alongside server-side functions and analytics. The same changelog documents the spend-approval rules for the Google Ads and Link shopping features Replit announced on its blog on Nov 12. Social discussion of the new items was minimal.
Sources
Mercor study: frontier agents miss risks buried in files far more than stated ones
OtherNewsX
Mercor published research on Nov 13 testing five frontier models on twelve simulated law, banking and consulting tasks where an agent is asked to send, delete or otherwise act. When the risk was spelled out in the prompt, frontier models acted unsafely about 1% of the time, but when the same risk sat only in workspace documents the rate jumped to 60%, and sub-frontier models did far worse. In one legal example, a model deleted files under a preservation letter after noting a possible spoliation concern without opening the letter. Each task also had a benign control version to check for over-refusal, and the authors argue implicit-risk judgment is measurable and improvable.
Sources
Developers debate Perplexity Decider v1.3's gains as the open-weight classifier hits HN
DiscussionHNNewsXBluesky
Perplexity's Decider v1.3 and Flash v1.3 decision models, released with a span-extracting locate mode and Apache-licensed weights, reached the Hacker News front page on Nov 13 with over 100 points. Trade coverage stressed the low per-request pricing and that all benchmark wins over TypeSafe's Jev are self-reported. Commenters were split: some found it strong and cheap on their own tests, while others said the main model barely moved versus v1.2 on a public decision index and questioned the flash-tier trend. Perplexity's CEO separately said its decision models now serve about a billion requests a month on OpenRouter.
Sources
1 more source
Cognition launches Devin Apps: private team web apps hosted on Devin's cloud machine
LaunchXHNNews
Cognition introduced Devin Apps, letting teams describe an internal tool and have Devin build and host it as a private web app on Devin's own cloud computer. Apps inherit the session's MCPs, integrations and codebase, get a stable sign-in-only URL shareable to people, an org or the whole enterprise, and keep prior versions serving if a new deploy fails. They run on a single machine that hibernates via memory snapshots and bills only while awake, so Cognition pitches them for dashboards, internal tools and staging copies rather than high-traffic production. Reaction on X was enthusiastic, with users praising the pace of shipping, while several asked how apps are isolated from workspace secrets and whether they can reach systems behind company auth.
Sources
1 more source
More serverless degradations hit GLM 5.3 Flash, DeepSeek V4.1 Flash and Kimi K3 Fast
ComplaintXBluesky
Fireworks' status feed logged another run of serverless incidents on Nov 13, including degraded performance for GLM 5.3 Flash, DeepSeek V4.1 Flash and Kimi K3 Fast. One GLM 5.3 Flash incident was marked resolved after about 11 minutes. The outage tracker outages.live ranked Fireworks first for incident count that UTC day with eight, ahead of OVHcloud and Cloudflare. The pattern extends the previous day's degradations and was shared on X and Bluesky by monitoring accounts rather than drawing much direct user commentary.
Sources
1 more source