Highlights
What happened at the companies we cover, by day, newest first. Every item links to its sources.
Showing 1–15 of 15 highlights
Sunday, Nov 8 15 stories
Investors keep debating Nscale's Microsoft exposure as Microsoft trims neocloud partners
DiscussionXNewsBluesky
Discussion on X through Nov 8 kept circling reports from earlier in the week that Microsoft passed on expanding its Nscale contract in Texas and, per CNBC and Bloomberg on Nov 6, plans to shrink its roster of neocloud suppliers in favor of its own data centers. Posters highlighted that most of Nscale's roughly $103B contracted backlog disclosed in its September IPO filing sits with Microsoft and Anthropic, with only a small slice already active. Skeptics framed Nscale as the clearest casualty of the review, while defenders argued the earlier UK exit stemmed from grid-connection delays rather than Nscale itself. A report also noted Nscale edited its Texas campus announcement to drop wording about a Microsoft expansion option while keeping the 700MW site capacity claim.
Sources
1 more source
Cursor users prepare for OpenAI models leaving the editor on Nov 12
DiscussionX
With OpenAI's proposed Nov 12 end to direct model access in Cursor four days away, developers on X spent the day passing around migration advice. OpenAI announced the wind-down in late August after SpaceX bought Cursor's parent Anysphere. Posters noted that bringing your own OpenAI key only works for local Chat and Agent, not Tab, Auto or cloud agents, and that the Codex extension is another way to keep OpenAI models inside Cursor. Several also repeated Cursor's estimate that OpenAI is only about 5% of its traffic. Reactions split: some said losing OpenAI pushed users toward Anthropic models, while others complained that top-tier plans now lean on Grok and Composer models they consider weaker.
Sources
1 more source
Clay details the 17 automated GTM workflows its own GTM engineering team runs
DiscussionXOther
A Nov 8 Growth Unhinged deep dive, written with Clay's head of GTM engineering and disclosed as Clay-sponsored, lays out 17 workflows Clay runs on its own platform. They span account fit and propensity scoring written back to Salesforce, territory assignment, signal alerts in Slack, pre-meeting research briefs, live org charts built by an account agent, follow-ups drafted in each rep's own voice from Snowflake data, pre-filled CRM updates and churn-risk health scoring. The recurring design choice is human-in-the-loop: drafts land in inboxes or forms for reps to approve rather than sending automatically. The piece is circulating among GTM-engineering practitioners on X, with readers pointing to the approval-first pattern.
Sources
Lovable apps can now take PayPal checkouts, captures and refunds via a new connector
LaunchOther
Lovable's Nov 8 changelog adds a PayPal connector for apps built on the platform. Builders can now create checkouts, capture payment once a buyer approves an order, issue refunds and read order details in an existing PayPal account. Lovable pitches it for shop checkouts, donation pages, ticket sales and booking deposits, and it can be added from Connectors or by asking the agent in a project. It joins the Mollie connector added Nov 3 as part of a steady expansion of payment options for AI-built apps; there was little social reaction on launch day.
Sources
Builders weigh Perplexity Teammates' design against its $200 Max-only gate
DiscussionXBlueskyNews
Discussion of Perplexity's always-on Teammates agents kept building across X and Bluesky four days after launch. A Nov 8 IT Brief rundown added detail on how they work: each teammate runs on its own isolated computer and can spawn parallel sub-agents, it connects to Slack, GitHub, Linear and Gmail through plugins, and it defaults to Claude Opus 5.5 while drawing from Max credits. Many posters liked the shift from answering prompts to owning recurring jobs with human review before work ships. The most common objection was access, since it is a web-only research preview for individual $200-a-month Max subscribers and isn't on Pro or by default on Enterprise plans.
Sources
Creators rush FLUX 3 Max video tests as fal's free window closes
PraiseXBluesky
With fal's no-cost FLUX 3 Max generation window ending Nov 9, creators on X kept posting text-to-video and image-to-video clips made with Black Forest Labs' video model. Many users called the output among the most realistic available, with one tester reporting a 30-second sequence rendered far faster than a Veo 3.1 run, and others sharing prompt tricks like scripted hard cuts to get multiple shots in a 15-second clip. Some results were described as mixed after a single try, and resellers began offering discounted access after the free period. Separately, a few users asked again whether FLUX 3 weights would be released, echoing earlier open-weights pressure.
Sources
1 more source
Public PoC for Hugging Face datasets pickle RCE spreads with no patched release
NewsNewsXBluesky
A security outlet report dated Nov 9 brought fresh attention to CVE-2026-57845, an insecure-deserialization flaw in Hugging Face's datasets library (versions up to 5.1.0) that lets a crafted pickle file run Python code when loaded through the deprecated pandas builder. VulnCheck disclosed it on Oct 23 with a CVSS 4.0 score of 8.7, and the researcher's write-up and working proof-of-concept are public. The proposed fix, which makes the pandas builder error out before opening any file, is still an unmerged pull request, so no fixed version exists yet; the report cites roughly 38M monthly downloads and no known in-the-wild exploitation. The report was reshared by security accounts on X and Bluesky, with advice to avoid that builder for untrusted data and use Parquet or Arrow instead.
Sources
1 more source
Users report Instinct clearing inboxes, but note connector limits and moat doubts
DiscussionX
Early users kept posting what they hand to Instinct. One said the agent worked through more than 200 emails and sent follow-ups to about 30 contacts after a conference, and others said they were excited about it or shared casual weekend use. A reviewer who tried it overnight found it useful but limited, since users can't add their own connectors yet, and still preferred rival agents. Separately, a widely shared post argued that no personal agent, Instinct included, has a durable moat, because agents remove the switching costs that usually create network effects.
Sources
Agentic-Siri report sparks X debate over whether Apple should buy Instinct
DiscussionX
After a Nov 8 Bloomberg report that Apple wants to make Siri more agent-like, several X posters argued about whether Apple should simply acquire Instinct, with ~$30B floated as a price. This is commentator speculation only; no report says Apple is in talks with Instinct. Backers argued that a deal would be small next to Apple's market cap and could save it 18 to 24 months of work. Skeptics said Instinct isn't worth that much, that Apple has invested little in AI, or that Apple will eventually win the OS-level agent race on its own and leave Instinct investors exposed.
Sources
Replit mobile app chat fails for about 20 minutes before fix ships
OtherBlueskyOther
Replit's status page logged an incident on the evening of Nov 8 PT saying chat was not working in its mobile app, classed as a major-impact issue for that component. The company posted that it was investigating shortly after 5 PM PT and reported a fix in place with mobile chat working again about twenty minutes later, then moved to monitoring. Several automated outage trackers on Bluesky relayed the alert, but there was little user complaint volume and no sign the web workspace or Agent were affected. It follows a Nov 1 Agent and workspace outage and a Nov 2 chat access incident, so reliability remains a recurring watch item.
Sources
Sandbox rival boat's comparison pages spark pushback over how they portray Modal
DiscussionXOther
Sandbox startup boat publishes comparison pages ranking itself against Modal, saying Modal sandboxes can't handle a long-running 'software factory' workload without a built-in desktop, persistent disk and Docker, and citing a third-party CPU benchmark where boat beat Modal's gVisor and VM runtimes. A widely shared X post on Sunday called the pages disingenuous, arguing Modal handles persistence via filesystem snapshots and that installing a desktop per image is a choice, not a gap, and that the benchmark compares unlike setups. A boat founder replied that Modal sees the startup as a threat, while others debated whether Modal is mainly a GPU/inference cloud with sandboxes as a side product. The page itself concedes Modal wins on GPUs, regions, egress controls and large bursts of short jobs.
Sources
ACR plenary study ranks OpenEvidence first of five AI tools on rheumatology guideline accuracy
NewsNewsXBluesky
A study presented at the ACR Convergence 2026 Plenary I session on Nov 8 ran 60 standardized rheumatology scenarios built from ACR and EULAR guidance through ChatGPT, Gemini, Claude, Perplexity and OpenEvidence. OpenEvidence scored highest on guideline concordance at about 91% and had the lowest hallucination rate near 4%, versus up to roughly 18% for the weakest platforms, while ChatGPT gave the most complete answers and responded fastest. The authors stressed that every platform made clinically meaningful errors such as wrong dosing or outdated drug choices, so outputs should be treated as drafts that need clinician checks. Rheumatologists amplified the result on X and Bluesky, and OpenEvidence said attendees visited its booth to discuss it.
Sources
1 more source
Vercel CLI auto-upload backlash; Rauch says bare `vercel` will stop deploying
DiscussionX
A developer trying Vercel Sandboxes reported that running the bare `vercel` command after login and project linking immediately uploaded the whole working directory, and noted code-training/sharing settings appeared enabled by default. CEO Guillermo Rauch replied that deploying should belong to `vercel deploy` and said the team will change the bare command, acknowledging it is a breaking habit for existing users. The thread kept going on Nov 8 PT, when the original poster added that, per his own decompile, the upload did not honor .gitignore and picked up .env files, and others asked whether anyone who ran it should rotate keys. Reaction is mostly critical of the default, with the CEO's quick commitment seen as the fix.
Sources
1 more source
Suno defends its artist-name prompt blocks as Elvis and Eminem suits draw wider pickup
NewsNewsXBluesky
Suno is now fighting two more lawsuits, one from rights holders of Elvis Presley's catalog and one from Eminem's publisher Eight Mile Style, which say the generator outputs songs and lyrics that imitate those artists. The Presley complaint also leans on Tennessee's ELVIS Act, a state law targeting AI cloning of a performer's voice. In response, Suno said it bars prompts that name specific artists or songs and uses outside vendors to screen uploaded audio and lyrics for misuse, and that it will keep adding safeguards. The Verge and Engadget picked the story up on Nov 8, placing it alongside existing suits from Sony, UMG and others; on X and Bluesky the reaction mixed neutral link-sharing with renewed creator backlash against AI music.
Sources
Developers split on SWE-2 Flash: free Devin access praised, benchmarks and closed weights doubted
DiscussionXBluesky
Four days after Cognition released SWE-2 Flash in Devin Desktop and CLI, developers on X kept trading tips for using it and the base SWE-2 model at no cost through the end of November. Several users called the free, fast SWE-2 variants the best budget coding setup available, often pairing them with a separate frontier model as orchestrator. Skeptics pointed to Cognition's own benchmark table, where the cheaper model trails rivals on most tests, and one independent eval scored it as unsuitable for serious development. Others questioned the claim of beating American open-weight models while Cognition itself keeps the fine-tuned weights private behind a planned partner program.
Sources
2 more sources