ZeroNoise Logo zeronoise
Post
SaaStr says agents now pick its software as HubSpot refounds and the fight over Nvidia's ouster continues
•
6 min read
• 1623 docs
SaaStr's operator report on HubSpot's "Customer Harness" pivot shows how agents can drop a vendor without anyone cancelling. Also covered: the continuing dispute over Jensen Huang's exit, model weights that point to Blackwell training at four Chinese labs, Meta's lessons from launching Muse, and research and security signals investors should weigh.

SaaStr: agents now choose the software, and they can drop vendors without cancelling

Brian Halligan says he and Dharmesh Shah are refounding HubSpot as the "Customer Harness Company" . SaaStr's write-up describes what that looks like at a company with 3 humans and 20+ agents in production :

  • Nobody on the team logs into the CRM. Agents read and write it through MCP and the API, and it stays the system of record .
  • Marketing automation is gone. In 60 days, marketing to a 450,000+ list moved to agents. They rebuild templates for each send and write emails personalized to each recipient. Consent and deliverability are bought as API services .
  • Agents are picked one job at a time, on evals, not bundled from the platform vendor. As SaaStr puts it, "owning the data can just mean someone else's agent calls your API" .
  • Churn without a cancellation. A core vendor went down for 48 hours, the agent routed around it, and the team never went back. "Usage just falls to zero." The reverse also holds: a winning agent gets sticky, because every switch means new evals and new context .

For SaaS diligence, this hits seat-based pricing and login-based health scores directly . It is one operator's report, but it describes a way to churn that renewal data would not catch.

Jensen Huang's exit: the fight is now about the reporting

Tim Sweeney says the articles have "backtracked" to "Jensen did lawful things that powerful people didn't like." He argues the original claims appear to come from a single competing Big Tech lobbyist, and he wants due process instead of press takedowns . Jason Calacanis guesses that White House China hawks leaked the memo , and calls it absurd that Huang would knowingly break export controls . Ed Zitron asserts that Huang "obviously helped a massive illegal smuggling operation with Megaspeed" . All of these are opinions. None is a finding.

On the economics, Bindu Reddy calls Nvidia's 80% margins unsustainable and says financing neoclouds "propped up" Nvidia's own demand . She predicts AI will "fizzle" rather than burst, with private valuations falling 10–100x . @goodwin_ml, quoted by Harry Stebbings, makes the opposite case: GPUs that are powered on are fully utilized, there are "no dark GPU fleets" , and frontier model providers are building first-party chips to protect their margins .

China's open models: Blackwell fingerprints, and a Huawei-only release

An analysis finds numerical traces in six published models from DeepSeek, Moonshot, MiniMax and Ant Group that match Blackwell chips, where those labs' earlier models matched Hopper. The method was right on 24 of 24 known cases . The report itself says it cannot show where the chips were or whether export controls were broken, and Blackwell can be rented outside China . The newest Qwen, GLM, Tencent, Xiaomi and StepFun models still read as Hopper . With chip controls under scrutiny, investors should treat this as evidence, not proof.

Meituan released LongCat-2.5 under the MIT license. It has 1.6T total parameters with about 48B active. On Terminal Bench v2.1 it scores 88.4 against Claude Opus 5's 89.1, but on v4.0 it trails 38.5 to 53.9. The model card documents deployment only on Huawei NPUs, with no Nvidia path yet .

Meta Muse: what launching a consumer agent took

On Lenny's Podcast, Mona Sarantakos said Muse came out of a personal-agent experiment she ran over a weekend . It took about seven months to launch, while some competing assistants took a couple of months . Most of that time went to reliability, safety and security. Adding a browser opened new attack vectors, and the team's internal rule was "trust is a must" .

Shopping turned out to be a top use case and home automation underperformed, so the team rebuilt its evals around shopping . Agent swarms produced multi-turn evaluation data that Meta did not have outside coding . The system prompt and behavioral spec must be written by hand . Activation is still "the number one problem": word of mouth works best, and later users need different onboarding than early adopters .

Research worth watching

  • Editing weights directly (MIT). GPT-6 Astra edited about 0.016% of a small Qwen model's weights, with no training, and its math score went from 7.9% to 48.1% . The caveats matter. Most of the gain came from fixing a repetition habit. Answers hitting the length limit fell from 93% to 33.1%. There was no established gain on Gemma, and independent reproduction of the headline result is pending .
  • Gödel's notebooks. Claude produced a 650-page explanation of Gödel's notebooks in a couple of days, though it had to be checked page by page. It found Gödel was "one step away" from a proof of the independence of the axiom of choice .
  • LeCun. He says superintelligence will not arrive within two years and that AI remains "terrible" at understanding the physical world. He is working on this at a new company the transcript renders as "Emmy Labs" .
  • Andrew Ng. He says he will no longer hire engineers who only implement specs, because user feedback, not code, is now the bottleneck . AI projects often stall around 80% until teams switch to golden datasets and rigorous evals . Skill with coding agents does not yet appear in job descriptions .
  • Hassabis. He calls AI that improves itself "an extremely useful technique" that "carries the most risk," and says the race toward it is outpacing safeguards .

Security and diligence

According to Cursor, a stolen Fireworks AI key was embedded in a trojanized Claude Code VS Code extension. The failure was in the supply chain, not the model . The post points to static keys, no per-call endpoint verification, and no way to detect a credential swapped mid-session . That points to a gap in tools for securing agent credentials.

Investing in AI argues that vibe coding creates "Recovery Debt": hours saved writing code are borrowed against future time to recover from failures . The author would watch observability, testing, reliability and incident-response companies. In diligence, ask whether mean time to recovery rises as the share of AI-generated code rises .

Briefly

  • An analysis by Hill estimates that an Anthropic IPO could make about 4,250 of its 5,300 employees millionaires. That assumes a $1.5T valuation, above the roughly $1T target . Clem Delangue answered that concentration of power and wealth is AI's biggest risk without much stronger open source .
  • Allie K. Miller notes that ChatGPT ads emphasize consumers, while most enterprises she speaks with "still lean Anthropic" .
  • Apple's health-AI study with OpenEvidence can enroll up to 200,000 US adults. Answers link to clinical sources, and Apple says study data will stay out of AI training .
SaaStr says agents now pick its software as HubSpot refounds and the fight over Nvidia's ouster continues
Lenny's Podcast
  • Muse emerged inside Meta after Mona’s weekend experiment with a personal agent connected to email, calendar, and home automation produced a useful family-week briefing from a short prompt; adapting the experience for consumers required safer operation, easier setup, and scalability. A major prelaunch focus was reliability, safety, and security: the team red-teamed the product, browser capability introduced new security vectors, and trust was an explicit launch gate.
  • Real-world use shifted product priorities: shopping became one of Muse’s top use cases, while home automation was less popular than expected, prompting the team to add shopping-focused evaluations. Mona sees potential for consumer agents to interact with merchants’ business agents, while arguing that agent-assisted decisions should still preserve browsing and inspiration.
  • Consumer activation remained difficult: Mona identified relatable word-of-mouth stories as the most effective way to show the product’s value; users also freeze on the first prompt, and later cohorts needed different onboarding to understand how Muse differs from other AI products.
  • Muse used agent swarms to analyze conversations and feedback and generate synthetic data for evaluations, helping address the lack of production data for long-running, multi-turn tasks outside coding. Its evaluations cover safety, capabilities, behavior, and use cases, with human review required for evaluation data and product specifications.
  • Mona’s stated vision is for agents to move beyond apps into ambient devices such as glasses or charms; she also wants agents to require less user babysitting and take autonomous actions more reliably as their capability and trust improve.
Meta’s Muse: How we launched the most successful consumer product this year | Mona Sarantakos
Andrew Ng
Profile
  • AI coding agents have shifted the software bottleneck toward user feedback and product judgment: Andrew Ng says implementation has accelerated beyond product managers’ ability to specify work, and argues engineers should help decide what to build rather than only implement a spec.
  • Ng says AI projects can advance quickly through prototyping but then plateau; moving toward higher reliability calls for golden datasets, rigorous evaluations, error analysis, and systematic tuning, with the right approach changing as a project matures.
  • His team analyzed more than 20,000 job descriptions, along with interviews and surveys; research postings emphasized machine-learning foundations most, followed by foundation-model skills and, less strongly, evaluation-driven development. Coding-agent skills were not yet appearing in job descriptions despite hiring managers’ agreement that effective use was needed.
  • Ng says AI is speeding research implementation and data analysis, while evaluation design, literature understanding, problem selection, and contextual judgment remain difficult. He sees physical-science automation using robotics, experiments, or simulation—and automation of AI and computer-science research—as promising, while broad full automation remains difficult.
BayLearn2026 05 AndrewNg HD 1080p
Yann LeCun
Profile
  • LeCun cautions that superintelligence is not coming in the next year or two; he describes current AI as useful for specific tasks but weak at understanding the physical world, a gap he says helps explain the absence of domestic robots and fully autonomous Level 5 cars. He identifies physical AI as a major challenge he is working on. He says he is working on it at a “new company,” rendered in the transcript as “Emmy Labs.”
  • LeCun expects AI to become widely used in research, including helping conduct experiments and potentially automating experimental processes in chemistry labs. He sees opportunity in scientists combining domain knowledge with enough AI expertise to build novel tools for discoveries in materials science, biology, medicine, and environmental protection.
Yann LeCun Unfiltered: The Godfather of AI on Life, Logic, and the Future
Fei-Fei Li
Profile
  • The explainer identifies the 2012 convergence of GPU compute, large datasets, and algorithms as a turning point for modern AI, and says the ImageNet project collected more than 15 million images. It contrasts AI’s need for millions of images to recognize cats with a child reportedly learning from 3–10 examples, illustrating the gap between data-intensive machine learning and human high-context learning.
  • For healthcare, it presents AI as augmenting clinicians: machines can contribute precision and rapid data analysis, while surgeons provide patient-specific judgment; AI can also analyze symptoms to support doctors’ assessment and treatment.
15 Why Modern AI Models Need Billions of Images While Human Bra
Paul Graham

Paul Graham argues that founder ability, rather than investment, is the limiting factor, and observes that decades spent solving hard problems can still be dismissed as work “anyone could have done.”

You work your ass off solving really hard problems day after day for decades, and this is what you get: anyone could have done it. As an …
Elad Gil

Elad Gil highlights Nvidia’s public-market compounding: its market cap increased almost 1,000× over 18 years, with $1,000 invested then worth close to $1 million today. The linked retrospective says Nvidia was worth about $5 billion in early 2009, versus Intel at about $79 billion, and recounts Jensen Huang demonstrating an Nvidia GPU whose computing performance he said equaled 54 times the world’s fastest supercomputer in 1993; the post frames that GPU as a chip now central to AI compute.

Almost 1000X increase in market cap in last 18 years as public company $1000 invested then would be close to $1M today [https://x.com/rdo… In early 2009, Nvidia was worth about $5 billion and Intel about $79 billion. That was the backdrop when Jensen Huang, then Nvidia’s CEO,…
@jason

Jason pushed back on Ed Zitron’s allegation that Jensen knowingly helped an illegal Megaspeed smuggling operation , saying it was absurd to think a leader of the world’s largest, rapidly growing company would knowingly violate export controls and calling Zitron’s stance a “bit.”

This is all posturing to try and force the board’s hand. Jensen obviously helped a massive illegal smuggling operation with Megaspeed and… Absurd to think Jensen, running the largest company in the world, that’s growing at an absurd rate, would ever knowingly break export con…
Allie K. Miller

Allie K. Miller says she will open paper mail with iMessage open to OpenAI Dots, and extend her mass-unsubscribe approach from email sales lists to brochures and letters; the post does not specify Dots’ function or report broader uptake.

Moving forward, I will always open my paper snail mail with my iMessage open to OpenAI Dots. I’ve already mass unsubscribed from email sa…
martin_casado

Martin Casado sarcastically pushed back on Matthew Stoller’s claim that SpaceX’s progress is the ordinary result of applying capital and engineering, and that Musk’s distinguishing skill is obtaining Wall Street capital, by comparing that reasoning to saying general relativity is the ordinary result of thinking hard.

… uncovering general relativity is the normal consequence of thinking really hard about something. [https://x.com/matthewstoller/status/2… What SpaceX did today is the normal consequence of applying capital and engineering to any industry. You get better at it. Musk isn't a g…
Machine Learning
  • Meta FAIR + NYU tested 200+ MLLMs trained from scratch on the same data, counted encoder pretraining against the same lifetime-compute budget, and evaluated on 16 VQA and 12 language benchmarks . In their tests, separately pretrained vision encoders at best tied encoders trained from scratch at matched compute; vision-only parameters, however, were needed, with encoder-free models trailing at every tested size from 0.6B to 6.4B and requiring about 3× the compute to catch up .
  • Their “in-fusion” design—vision-only weights in the LLM’s first eight layers—matched encoder-based results from 1.7B to 6.4B and reduced time to first token by about 13%; the authors also report the finding held when adding vision during midtraining of a pretrained LLM .
  • A discussion caveat is that the same-data comparison may not settle the value of scaling on unlabeled image/video data: a commenter argued that caption requirements limit CLIP’s usable data, while a reply reported comparable results from WebSSL, a pure-vision SSL method, and left its data-scaling potential open .
[Meta FAIR + NYU] Unbundling Multimodal Vision Encoding: training 200+ MLLMs to find encoder pretraining is optional but vision-only parameters are not [R] The elephant in the room that this paper kind of ignores is pure-vision SSL. MetaCLIP that they used as their main example of a vision en… We did try WebSSL, which is similar to DINOv2, as a pure-vision SSL method and it does comparable (Figure 2a). It would be interesting to…
Paul Graham

Paul Graham’s advice for young founders is to prioritize users over investor approval: users are the ultimate gatekeepers, and winning them can bring investors; investor enthusiasm is temporary if users dislike the product.

The most consistent mistake I see young would-be founders make is to care too much about investors' opinions and not enough about users' … Or if it's easier to understand this using the old paradigm, users are the ultimate gatekeepers. Get them and investors will follow. And …
Aravind Srinivas

Perplexity CEO Aravind Srinivas says users can build Teammates with any frontier model, which he claims no other product allows. A user says choosing the model and effort level for each task brought them back to Perplexity after a hiatus.

Teammates are awesome, and you can build them with any frontier model of your choice on Perplexity. No other product lets you do this. [h… Perplexity Teammates is really growing on me, really dig that I can select model and effort level. This has been one of my biggest challe…
Sriram Krishnan

Sriram Krishnan praised an unspecified SpaceX achievement as a testament to what humanity can accomplish through science and engineering, congratulating Elon Musk and the thousands at SpaceX involved; the post does not identify the achievement or provide operational details.

can't get over this really happened. what a testament to what mankind can accomplish and the magic of science and engineering. congratula…
Allie K. Miller

Allie K. Miller said she taught hundreds of business professionals to build an “AI factory” organized around system-wide goals, real-time context, execution, decisions, and feedback/action loops .

Last Thursday, I taught hundreds of business professionals how to build their own AI factory (system-wide goals, real-time context, execu…
Allie K. Miller

Allie K. Miller argues that proactive AI assistants should resolve conflicts according to user priorities rather than merely flagging them, and avoid incessant alerts that add stress . She advocates a user-settable “proactivity temperature” and anticipatory actions that add delight, such as booking a cancellable appointment around upcoming events .

Dots, Muse, Duet, and Instinct can induce more stress if you don’t prompt it otherwise. And you should because an AI that incessantly pin…
@jason

Jason’s best guess is that China hawks in the White House leaked an intelligence memo that the linked All-In clip says led to Jensen’s resignation; the clip frames the issue as an internal White House split over whether China should receive chips. This is speculation, not a confirmed attribution.

My best guess, is that the China hawks in the White House are the leakers The call came from inside the (white) house! [https://x.com/the… Jason on the leaked intelligence memo that led to Jensen’s resignation: “The leak is coming from within the administration” “Who benefits…
Elizabeth Yin 💛

Elizabeth Yin promoted Batter Up New York, a pitch-preparation event for New York founders scheduled for November 16, 2026; the post framed it as a last-chance opportunity to fine-tune pitches.

New York founders, last chance alert! Don't miss Batter Up New York on November 16, 2026, to fine-tune your pitch! Register now: [https:/…
sarah guo

Sarah Guo highlighted an Apple research study exploring AI use in health alongside OpenEvidence, linking to Apple’s announcement; the post does not provide study details.

Apple Health AI x [@OpenEvidence](https://x.com/OpenEvidence) research study! [https://www.apple.com/newsroom/2026/11/apple-launches-rese…
Nathan Benaich

Nathan Benaich plans to host Airstreet AI meetups in Munich and Zurich in February 2027, while traveling to the Munich Security Conference; the meetups will feature a State of AI update, talks on AI research and product building, and time for peer networking.

in feb ’27 i’m off to munich for the munich security conference, then zurich. i’ll host [@airstreet](https://x.com/airstreet) ai meetups,…
Brad Feld

Brad Feld said Google Drive search had become poor and that he rebuilt Drive on GitHub, linking to a write-up hosted at Adventures in Claude.

For a search company, Google has made search in Google Drive suck badly. So I rebuilt Drive on GitHub. [https://adventuresinclaude.ai/pos…