Cognition announces Devin Cloud CLI and SSH workflows
Today's top 20 insights for PM Builders, ranked by relevance from X, LinkedIn, YouTube, and Blogs.
Cognition announces Devin Cloud CLI and SSH workflows
#1 đť•Ź
Cognition announced Devin Cloud in Terminal and devin ssh, two new ways to use Devin’s computer. Users can create, steer, and resume Devin Cloud sessions via `/cloud` in the CLI, or SSH into Devin’s dedicated VM and use `/handoff` to return work to their own device.
#2 đť•Ź
Qwen-Image-2.1 has open weights and runs locally on DGX Spark, RTX GPUs, and DGX Station.
Also covered by: @Qwen
#3 đť•Ź
v0 announced that Grok 4.7 is available in v0, with 40% off until September 27th. Try it via the provided v0.app link.
Also covered by: @Cognition, @Guillermo Rauch
#4 đť•Ź
Santiago says running 2 agents on the same data causes coordination problems and shared a link to the Omnigraph GitHub repository.
#5 in
Dharmesh Shah announced that HubSpot integrated with OpenAI’s ChatGPT Ads, claiming it is the first CRM to do so. HubSpot is offering one free ChatGPT business seat for 12 months with the purchase of one seat, plus matching funds for up to $750 in ad spend.
#6 đť•Ź
LlamaIndex 🦙 released Grounded Confidence for Extract, adding field-by-field confidence scores to help apps automatically accept accurate results or route them for human review. The feature is available on Cost Effective, Agentic, and Agentic Plus.
#7 đť•Ź
Dharmesh Shah announced that he spent the weekend building a Granola integration for YouSpot, enabling note syncing, entity extraction, and context-graph linking, with the resulting information accessible through chat, MCP, and cloud agents. Meeting notes can also inform YouSpot’s daily brief emails. YouSpot is an AI-native CRM for one-person companies, with a free version and a $10/month plan.
#8 đť•Ź
DeepLearning.AI shared an analysis of Meta’s Muse agent, which assumes the model will be tricked and implements security at the operating-system level. The model never handles real credentials, tools run in isolated Linux containers, and an independent gatekeeper verifies outbound calls.
#9 đť•Ź
Aravind Srinivas announced that Seedance and MiniMax H3 are now available on Perplexity Computer for detailed video-generation workflows, particularly web apps and creative materials that combine several tools with a video model.
#10 đť•Ź
Ali Ghodsi shared Superhuman’s blog post about scaling GEC inference, calling the company’s AI infrastructure scaling “super impressive.”
#11 đť•Ź
Garry Tan shared an example of Capy.ai executing an ambitious bug-fix wave on GBrain, demonstrating clear task delineation, automatic parallelization, and a clean GitHub pull-request and CI workflow.
#12 ▶️
An ex-OpenAI researcher just deleted language from the LLM...
Fireship
Jev, created by ex-OpenAI researcher Diogo Almeida and TypeSafe AI, is presented as a System 1 classifier that removes language output and returns strongly typed choice, score, or null responses with claimed 200Ă— faster speed, 400Ă— lower cost, free output tokens, and zero hallucinations.
- TypeSafe AI raised $40 million; Jev accepts a question plus unstructured-text context and guarantees schema matching for one of three output shapes: choice, score, or null.
- Jev uses RLCD (reinforcement learning for calibrated decisions) to return a calibrated confidence value, such as 60%; identical inputs can still produce different outputs because the model is not deterministic.
- OpenJeb reproduces the Jev interface by reading option probabilities from a frozen Qwen-4B model in a single forward pass, requires no new training, runs on an RTX 3090, and has a WebGPU browser demo.
Also covered by: @Harrison Chase
#13 đť•Ź
Logan Kilpatrick said AI product builders should spend >25% of their time creating benchmarks and getting model labs interested in them, calling it the easiest way to accelerate a company’s progress.
#14 đť•Ź
Guillermo Rauch described an agent reproducing, simulating, fixing, deploying, and verifying a mobile in-app browser rendering issue, including creating an ephemeral Vercel deployment for an iPhone simulator. He argued that agents can test and QA software with unmatched intensity, leading to unprecedented quality and performance.
#15 in
Claire Vo recapped a deeply technical episode with Warp CEO Zach Lloyd on running an AI software factory, from initiating work anywhere and managing tickets through verification to measuring human touch per pull request. They also discussed scoring and evaluations for failure modes and building factories that improve themselves.
#16 ▶️
How to Save Money Now with ChatGPT Finances (6 Real Use Cases) | Ethan Bloch
Peter Yang
Ethan Bloch explains using ChatGPT Finances—available with ChatGPT Plus and Pro subscriptions—to connect read-only bank and investment data through Plaid, identify avoidable spending and tax opportunities, optimize rewards and fund fees, and build interactive models for major financial decisions.
- ChatGPT Finances is included in the $20-per-month ChatGPT Plus subscription; it connects financial accounts via Plaid, does not receive routing or account numbers, cannot move money, and Plaid refreshes account data several times per day.
- A subscription-review prompt analyzes the prior 90 days for price increases, duplicate charges, refundable fees, and unusually expensive household bills; Ethan Bloch cited an Evernote increase from $129 to $249 and a hotel-points booking that ChatGPT identified as incorrectly billed and helped get reimbursed.
- Recurring Finance tasks include an automatic weekly spending/net-worth update, a monthly subscription-price and necessity review, a monthly joint spending-and-investment email, and an idle-cash monitor; the team shipped 47 product updates in the preceding seven days.
Also covered by: @Peter Yang
#17 đť•Ź
Harrison Chase announced that SemIf, an open-source alternative to the “decision model” jev, is based on qwen3.5 and is being hosted free through LangSmith Gateway for the next week.
#18 đź“ť OpenAI News
Advisory group on mathematics and artificial intelligence - OpenAI announces the formation of an advisory group focused on the intersection of mathematics and artificial intelligence to guide research and best practices. The group will provide expertise to help shape development and safety considerations for mathematically-intensive AI work.
Also covered by: @OpenAI
#19 đť•Ź
Andrej Karpathy commented that an unspecified subject occupies a point on the LLM Pareto-optimal curve: an under-invested regime with large latent demand for no thinking, single-token output, low latency, and acceptable intelligence amid the race toward higher intelligence.
#20 ▶️
$30M Writer: Never write AI slop again
Greg Isenberg
Nicolas Cole explains a “personal language model” built from a library of approved language, personality details, and original ideas, with AI used to repeat and remix that library rather than generate net-new thinking.
- Cole divides content into three tiers: commodity content, personality content, and original content; commodity content contains broadly shared ideas, while personality content adds attributable life details and original content comes from prolonged thinking within a specific topic.
- After a Quora answer went viral, Cole reused the sentence, “When I was 17 years old, I was one of the highest ranked World of Warcraft players in North America,” in hundreds of pieces over the following five to ten years.
- Cole defines voice through word choice, sentence length, and sentence structure; for one ghostwriting client, adding two or three source-reference sentences such as “according to the New York Times” or “according to Harvard Business Review” changed about 1% of a draft and made the client say it sounded like him.