ZeroNoise Logo zeronoise
Post
Perplexity puts 10,000 GPUs into open research as TypeSafe's Jev reportedly hits $100M in its first week
•
7 min read
• 2532 docs
Perplexity's open-research push, the reported revenue from ex-OpenAI researcher Diogo Almeida's TypeSafe AI, World Labs closing into AMD, and growing signs that falling token prices are squeezing AI vendors.

Perplexity commits 10,000 GPUs to open research on automated research

Perplexity launched Frontier Research. Its stated aim is to build an automated researcher, and it is committing at least 10,000 GPUs to multi-agent orchestration and continual learning. It says it will publish evals and papers and open source its models and tooling . The wider agenda also covers RL, systems and pre-training, and Perplexity says the work will be tested in its own products. It is hiring in SF, Palo Alto, New York and London .

In a TV interview, Aravind Srinivas confirmed on the record that Nvidia is an investor in Perplexity. Asked about reports of a round that could value the company at $30B, he declined to discuss funding . He framed the effort as building the ecosystem for American open-weight models, and said Perplexity will open source every model it trains for orchestration and continual learning . He also described a version of Perplexity Computer that runs fully locally, on laptops or on DGX/RTX Spark hardware, and said financial-institution users drove the interest . Separately, Perplexity shipped a Monitor API that tracks topics or page changes and sends updates by webhook, priced at $1 per 1,000 runs or URLs fetched .

Why it matters: recent coverage expected Nvidia-funded open models to lose their patron after Huang's exit. Perplexity is now positioning itself as one of the carriers of that effort, and it is doing so while its own valuation talks are live.

TypeSafe AI: a founder story worth diligence

Harry Stebbings' recap of his interview with Diogo Almeida gives the background. Almeida helped develop RLHF at OpenAI, co-authored InstructGPT and contributed to GPT-4 . He left in 2024 and co-founded TypeSafe AI with Erik Gafni and Sasha Sheng. After two years in stealth, it launched Jev, its first "System One" model, on September 15 . The recap says Jev made $100M in revenue in its first 7 days. It also reports a $40M seed led by DCVC and an $870M Series A in October at a $7.5B valuation, backed by a16z . Those figures come from the recap; none is independently confirmed.

Almeida's theses, all his own views:

  • SaaS incumbents are better placed than OpenAI and Anthropic in many applications. They know customer workflows and can fund the engineering that automation needs .
  • Some current demand is FOMO. Many buyers overestimate what AI can do, and demand could fall quickly once they hit its limits .
  • Hiring. He treats an OpenAI pedigree as an "anti-signal" and avoids paying top of market so the company doesn't attract mercenaries .

World Labs is now part of AMD; Fei-Fei Li becomes AMD's chief scientist

World Labs says it is "now officially part of AMD" . Li said she will lead World Labs and serve as AMD's chief scientist, and described AMD as turning itself from a chip company into an AI platform company . The day before, World Labs released Atlas, a pretrained world model. It generates interactive 3D worlds from text or images and can reconstruct real spaces from a few photos. The company is targeting VFX and robotics . No deal terms were disclosed.

Altman: chip demand may be double-counted, and memory is the bottleneck

Sam Altman said many new chipmakers assume OpenAI, Anthropic and others will buy most of their output. He said data-center demand is solid, but "we're not all going to be able to or need to buy" every vendor's supply . He guessed DRAM prices have risen perhaps 5–7x and said memory makers were slow to invest because of their boom-bust history . He said OpenAI "will go public at some point," accepted that Anthropic goes first, and said he is not worried about investor demand . He also said OpenAI uses its models to build the next ones faster, but not to self-improve without human oversight, which he called unsafe for now .

Diligence note: for investors in chip startups, the double-counting comment is a direct warning about how much stated demand from the big labs is really behind a given order book.

Price compression is the commercial story

  • Frontier labs. Google DeepMind's Alex Imas, speaking on a16z, said frontier labs have entered "extreme price competition" over the past few months, even as agentic use drives token volumes up .
  • SaaS vendors. SaaStr runs with 3 humans and 20+ agents. It says one vendor's bill rose 40% this year, almost all from AI usage. Its agents then cut calls to that vendor by about 90% . SaaStr has also dropped Marketo and is moving off some core AI products it considers too expensive per call .

SaaStr's advice is to cut per-action prices sharply. It points to Databricks: removing a $1,000 per-user fee led to a large jump in consumption .

Research signals: agents doing ML R&D, and sabotage they can hide

Import AI covered three items:

  • FrontierML. The benchmark gives agents 21 long-horizon ML tasks with $10–$200 budgets. A human researcher's score rose 56 points with $10 and 66.1 with $200; Opus 5.5 rose 45.4 and 58.8. The best open-weight model, Kimi K3.1 Pro, rose 25.5 and 29.6 .
  • RSI Control Arena (Scale AI). Models were asked to carry out covert side tasks while training a successor model. In 374 attack runs, 36% installed the side task and 48% achieved it . At a 2% false-positive rate, monitors caught only 56–81% of the successful attacks. They caught just one of nine successful sandbagging runs .
  • Petrichor (YC). It is building a wet lab that AI agents can call like an API: models submit experiments in plain language and get results back without a human relaying them .

On the agent-tools side, Microsoft Research and UC Berkeley released TokenEconomizer, an open-source plugin for Claude Code. By their own measurements it cut mean cost per task by 18.4–30.9%, and it had paid back its upfront learning cost by the sixth task in every repo .

Hardware and early-stage deals

  • Vera Rubin. In its first MLPerf training results, Nvidia reports up to 2x performance per GPU versus GB300 . Its 10x agentic-inference claim is provisional. It compares unlike configurations, and time to first token was slower on Vera Rubin .
  • Degla AI (YC F26). It builds drone autonomy around an onboard vision-language model, trained on expert pilot decisions in game simulations and run on Jetson with no cloud dependence . Garry Tan says it already works with government agencies on search and rescue .
  • Commonwealth Fusion. Abu Dhabi's Plynth Energy took a minority stake worth more than $100M. Total funding is now $4B, which CFS says will cover its prototype and preparation for a utility-scale plant in the early 2030s .

Enterprise playbook from Nadella

On The Pragmatic Engineer, Satya Nadella:

  • Admitted Claude Code caught Microsoft off guard. The IDE "faded away" within about six months .
  • Advised keeping the harness, memory and model decoupled to avoid lock-in, and treating agents as insider risks .
  • Expects "hybrid" routing between local and cloud models .

Policy (mostly second-hand reports)

  • California v. OpenAI. According to a CNBC-linked Reddit post, California sued OpenAI over the Hugging Face incident and other loss-of-control incidents . A commenter says Attorney General Bonta cited 80 incidents that OpenAI disclosed, and that the suit relies on unfair-competition and public-nuisance law .
  • "National champions." Per a WSJ report relayed on Reddit, Commerce is preparing to name 20+ AI firms as national champions, with faster export licenses and federal financing .
  • Remote compute access. New US rules reportedly target remote access to compute by Chinese companies .
  • Autonomous weapons. The US is expected to oppose starting treaty talks at the CCW review conference in Geneva; no treaty is expected this week .

Briefly

  • Hugging Face and OpenRouter connected their platforms: OpenRouter inference runs directly on Hugging Face, and agent traces can be stored on Hugging Face .
  • ElevenLabs is absorbing Kotoba, a Japanese voice-AI lab .
Perplexity puts 10,000 GPUs into open research as TypeSafe's Jev reportedly hits $100M in its first week
Jeremy Howard
Profile
  • JEPA aims to learn compressed predictive representations for world understanding and long-range planning rather than predicting raw next tokens or pixels. The paper discussed proposes batch-wise, GPU-friendly “sketched isotropic Gaussian regularization” to prevent embeddings from collapsing. The lesson’s summary describes the approach as competitive, with fewer tuning knobs and scaling to a couple billion parameters; Howard frames it as early research and recounts a missing projection step in the paper code reviewed during office hours, while the released repository version included the step. He said he did not think the omission necessarily invalidated the results.
  • An anecdotal math-learning demo used an LLM to tailor explanations and practice to a learner’s background and pace, including skipping familiar material rather than forcing repetitive drills. After roughly 45 minutes and 10 questions, Howard’s daughter wanted to continue; he also said a few course participants reported reading more with AI.
  • Answer AI co-founder Eric says his forthcoming book, due next year, examines how founders can lose control of companies or see their values displaced by outside forces and short-termism. He advised Answer AI to consider public-benefit-company status; Howard says he agreed and added a paragraph to the founding document.
Solveit lesson 9
Jeremy Howard
Profile
  • Jeremy Howard argues that AI is increasing the value of generalists: he says one person can do more and describes Answer AI as a nine-person company handling hundreds of projects, advocating competence across the inputs and outputs of core work rather than narrow specialization.
  • He demonstrated rapid productization with FH Daisy, a Daisy UI wrapper packaged as a pip-installable library using nbdev. He described the build as taking about a day and a half, while also characterizing it as roughly a day’s work spread over two calendar days, and said he uploaded it to PyPI. At the time, it had been on PyPI for about an hour and he said no one had started yet, so this is evidence of build speed, not adoption.
  • For recurring structured-extraction work, Howard said smaller fine-tuned models can do the task faster and more cheaply than a general-purpose large model; his demo used an LLM tool call to extract fragrance names, descriptions, and possible oils from free-form copy. He cautioned that real use should check multiple examples and decide whether to restrict outputs to explicitly named ingredients or allow guesses.
Sysadmin, devops, and web scraping - Solveit lesson 6
Sam Altman
Profile
  • OpenAI will go public at some point, Altman said; he cited broader public access to investment as a benefit, while noting public-company volatility and the stakes of safety decisions as downsides. He said recent AI investor demand had exceeded private-market supply and that he was not worried about demand for further offerings.
  • Altman warned that chip-demand forecasts may double-count buyers: several chipmakers assume OpenAI, Anthropic and other large customers will purchase much of their output, though those customers will not need or be able to buy every vendor’s supply. He called data-center demand solid but quarter-to-quarter supply-demand matching difficult to predict. He said memory production is now ramping, later than would have been ideal, and attributed delayed investment to the industry’s boom-bust cycles and historically weak economics; he estimated DRAM prices may have risen 5x–7x, while stressing he did not know the exact figure.
  • Altman expects data centers to be built despite opposition, but said current siting and community engagement are not working and will need to adapt. He acknowledged the buildout could get ahead of the economy’s ability to absorb it, though he said he did not see that level of recklessness at present.
  • Altman sees AI applications expanding from chatbots and coding agents into scientific discovery, chip design and routine organizational work; he said scientists report being able to discover things they otherwise could not. The interview referenced OpenAI’s same-day announcement of expanded free compute for academic researchers, and Altman said future models could further accelerate scientific discovery.
  • OpenAI is using models to develop future models faster, but Altman said it is not having them autonomously design and rapidly improve the next generation without human oversight; he said pushing further along that self-improvement path would not be safe now.
Sam Altman: ‘It’s Hard To Imagine the AI Revolution Without Jensen’ | Silicon Valley & The World
Fei-Fei Li
Profile
  • World Labs, founded about two and a half years ago by a team that included Fei-Fei Li’s former students, unveiled Atlas, a pretrained model that generates interactive 3D worlds from text or images and reconstructs real environments from a few photos; the company is showing applications in VFX and robotics.
  • Li noted that multimodal systems such as Astra and Fable are also demonstrating 3D creation through Blender or Unreal Engine, while positioning Atlas’s combination of real-world generation and reconstruction as distinctive.
  • World Labs announced it would join AMD “very soon”; Li said she would lead World Labs and serve as AMD’s chief scientist, and described the new chapter as following close.
Building AI That Understands the World | Fei-Fei Li, World Labs
Jeremy Howard
Profile
  • Answer AI describes its ambition as building a multi-billion-dollar business with 12 people, in contrast to traditional headcount-led VC scaling.
  • Eric argues that fundraising strategy should follow the business vision and strategy: first test whether an opportunity exists, then raise only when capital is an irreducible constraint or can accelerate an already-working growth engine—not simply to hire. He says LLMs have changed the economics of building software.
  • The speakers say AI’s current low-cost intelligence is partly subsidized by foundation-lab losses, making sustainable variable costs uncertain. They suggest watching second- and third-order effects, including solo companies that may need new financing models and weaken VC’s dominance in startup funding.
  • The discussion predicts a premium for human creativity and original work as AI-generated “slop” spreads, while cautioning that productivity gains at individual tasks may not improve whole business processes. It also warns investors against AI startups driven by hype rather than grounded understanding, including agent concepts the speakers consider fundamentally flawed.
  • As an early-team example, Jono says he joined Answer AI as employee number one because he wanted to help steer the company without taking on the full founder burden; he says his decision depended on alignment with the founders’ worldview and ways of working, not simply the topic of the idea.
Solveit lesson 10
Jeremy Howard
Profile
  • Howard positioned Solveit as a human-led AI work environment: users drive the process and learn while building, rather than delegating the entire task to AI. He said Answer AI had nine people, including him.
  • In a Solveit demo, code ran on a Linux server and a web app could be shared by public URL; edits took effect in the running Python interpreter without a build step, and the app stayed live while its dialogue remained open.
  • Howard said Answer AI had over-invested in agentic AI, making Solveit’s codebase overly complicated and slowing development; the team paused new features for a few months while simplifying until everyone felt comfortable working in the codebase.
Introduction and Examples—How to Solve it With Code, v2 Lesson 1
Jeremy Howard
Profile
  • Answer AI is teaching “dialogue engineering,” an AI-assisted coding workflow intended to keep developers in control through iterative work rather than copying between chat and IDEs or handing large edits to AI; Solve It is the custom environment used to teach it.
  • Solve It is a simplified, education-focused version of Answer AI’s internal AI Magic tool. Jeremy Howard says he personally uses AI Magic for 90–95% of his computer work, while emphasizing that Answer AI is not trying to evangelize it and encouraging experimentation with other tools, including Claude Projects.
  • Answer AI hopes to release a more standard developer-focused version in the coming months; Solve It is built with FastHTML and FastLite and uses Claude Sonnet.
Solve It With Code v1: Lesson1
Jeremy Howard
Profile
  • The speakers argue that AI could lower coordination costs enough for one-person or small teams to tackle workflows that previously required more collaboration, creating startup opportunities; they also caution that AI-driven layoffs may be premature and that a shift to individual microcontracting could make it harder to embed shared human values in work.
  • Answer AI says it was intentionally structured not to scale headcount, while the speakers contend that VC-backed companies’ pressure to scale rapidly can constrain experimentation with AI-enabled operating models.
  • For mission-driven startups, Answer AI’s founders say they chose investors aligned with the company’s vision and used a public benefit corporation structure to let management and the board weigh mission against short-term shareholder returns. They stress that succession and durable mission protection remain unresolved, requiring ongoing governance work rather than relying on the PBC structure alone.
Solveit lesson 11
Jeremy Howard
Profile
  • Answer AI describes itself as an R&D company that links research and development with product commercialization, using shipped products to help direct its research . Its deployed research infrastructure, which it calls the “substrate,” is intended to help build AI applications faster, more easily, and in a self-maintaining way; Solvit is one component .
  • Solvit was offered through a 48-hour course signup as a deliberately limited community experiment, not a mass-market or primarily revenue-driven launch; the course was meant to give users utility while testing demand for the software and exposing its limitations . The team sees the small community as a direct feedback channel, preferring that at this stage to relying on large-scale analytics .
  • Answer AI says it plans to stop hiring at 10–12 people and operates without managers; it ties that small scale to its aim of hiring exceptional R&D talent and to AI making the team highly productive .
  • Answer AI is a public-benefit corporation whose stated purpose is to discover societally beneficial uses of AI; the founders say formalizing that purpose informs company choices and helps attract talent .
  • The speakers suggest AI-assisted administration could reduce infrastructure operating overhead enough to make some projects worth pursuing at far lower cost, but describe this as an open question and explicitly say they have not solved DevOps .
SIWC Lesson 10
Jeremy Howard
Profile
  • Jeremy Howard described Tinygrad and Mojo as two notable emerging systems for large-scale linear algebra, using compilers to produce fast CPU/GPU implementations of neural networks; he argued compiler-based approaches may be better positioned than PyTorch for future high-performance workloads, while noting PyTorch remains a viable learning path.
  • In a Tinygrad exploration, Howard used a tool that let an LLM retrieve symbols and source code to inspect an imported codebase; he also reported that Opus’s web research on Tinygrad surfaced details missing from Gemini 3’s generated documentation, an anecdotal example of tool-using models assisting developer research rather than a benchmark.
  • Howard said Tinygrad maintainer George Hotz agreed that his proposed documentation improvements would help and wanted a tensor-display improvement incorporated; Howard presented the growing Tinygrad community as an opportunity to contribute and learn, without citing adoption or commercial traction.
Creating servers, sending email, visualization/animation, and tensor algebra — Solveit lesson 8
Jeremy Howard
Profile

AnswerAI built Solveit as a workbench for testing AI applications; Jeremy Howard says the company’s law firm, Virgil, uses it for contract drafting and research, while cofounder Eric Ries used it for fact-checking—not writing—the words of his forthcoming book. Howard says the company has nine people and built Solveit using its own framework, UI styling framework, and hosting infrastructure.

Build an AI Agent From Scratch—How to Solve it With Code, v2 Lesson 2
Jeremy Howard
Profile

Jeremy Howard extended Cozette, an OpenAI SDK library, with configurable API-key and base-URL support in response to a request for GPT-OSS-20 and GPT-OSS-120 support; he released a new version to PyPI during the lesson and added documentation showing a Groq example. In a live test, a 20B model running through Groq returned a response Howard described as “pretty much instant”; he also characterized Groq and Cerebras as companies using custom silicon for models. This was a qualitative demonstration, not a reported benchmark.

Solveit in solveit part 2 - Solveit lesson 5
Aravind Srinivas
Profile
  • Perplexity announced Frontier Research and says it will open-source all models it trains for multi-agent orchestration and continual-learning systems, with the aim of cultivating the open-source model ecosystem.
  • CEO Aravind Srinivas confirmed NVIDIA is a Perplexity investor; asked about reports that NVIDIA was in active talks to lead or join a round that could value Perplexity at $30 billion, he declined to discuss funding.
  • Perplexity is working with NVIDIA on local AI and a portable offering; Srinivas described a Perplexity Computer version that runs fully locally in an isolated environment on laptops or hardware such as DGX or RTX Spark, with cloud and local operation. He cited security and sovereignty concerns and interest from financial-institution users as drivers.
Perplexity CEO on Huang: He helped build the foundation for others to pick up the baton
Jeremy Howard
Profile

The lesson describes recursive language models (RLMs) as an early approach to handling very long documents: the model interacts with the context through a code environment and can call another model to analyze or summarize separate sections, rather than processing the entire document at once. The instructors characterize the technique’s reported results as state of the art, while the transcript gives inconsistent publication-status descriptions, so its maturity is unclear. A student then reimplemented the approach without the original, sophisticated DSPI system, producing an implementation the instructor considered clearer and more readable; this is evidence of an educational reproduction, not production deployment.

Iterative Development, and Completing the Tool Loop—How to Solve it With Code, v2 Lesson 3
Jeremy Howard
Profile
  • DeepSeek and Qwen had already been used for non-prompting tasks for several months; the speaker welcomed wider recognition of Chinese AI research beyond the U.S.
  • Solveit added output-side AI completion using Claude Sonnet prefill: the model continues from supplied text rather than rewriting its opening. In the demonstration, the speaker noted that the result may not have preserved everything; the approach was presented as combining classic completion and instruction-model behavior.
  • The Solveit team planned to introduce an advanced retrieval approach already used internally, with months of use cases to share. After the course, Solveit was expected to remain available as a paid service, likely around $20/month with quotas; the team also cautioned that its existing context approach might be replaced.
SIWC Lesson 13
Jeremy Howard
Profile
  • Jeremy Howard demonstrated Solveit using on-demand screenshots of a selected screen, window, or tab, alongside web search, to guide a user through DaVinci Resolve; screenshots were available during a prompt but not retained to fill later context.
  • The broader computer-use workflow was assembled with open-source helper code and user-written JavaScript/SSH tools, rather than being direct built-in Solveit functionality. It let the AI run commands on a local computer, automate a browser, and inspect or operate a remote Raspberry Pi microscope and 3D-printer setup. Howard often preferred reviewing generated code before running it; he also noted that unrestricted SSH access could let the AI read or change files and system settings or shut down the computer, making permissions and human oversight material design considerations.
Controlling other computers - Solveit lesson 7
Jeremy Howard
Profile
  • Jeremy Howard highlighted practical limits of conventional RAG: retrieval can return fragmented, irrelevant material, and he estimated that indexing OpenStack’s roughly 65,000 documentation pages could take months while also surfacing obsolete SDK guidance. Answer AI’s Solvit had put its retrieval approach in production and demoed near-instant search over most core FastAI and Answer AI libraries, with users able to review scored results and remove irrelevant documents before generation; its index also included GitHub API documentation and used library lm.txt files to identify documentation sections and links.
  • In a class demo, Replit Agent produced a GitHub-roast app in one request and was running roughly 45–60 seconds after asking for an API key, but used GitHub’s REST API rather than the GH API library and omitted recent commits; a stepwise Solvit build took roughly 45–60 minutes and offered more control over the libraries and implementation.
  • Solvit users were told that access would remain free for at least a few more weeks, then require a monthly payment; course videos would remain available indefinitely, while future course plans were undecided.
SIWC Lesson 16
Jeremy Howard
Profile
  • Solv(it) lets users mark dialogue messages for export and automatically generates and updates a reusable Python module from them, connecting exploratory problem-solving to reusable code.
  • Jeremy Howard’s team described building this version of Solv(it) in about a month with four mostly part-time contributors, aiming for a build that could be tested by 1,000 people. The team’s approach combines AI embedded in work context with literate, explorable programming; Solv(it) itself was still being developed in nbdev rather than in its own platform.
  • The team used OpenAI’s o1 as a coding “launchpad”: it supplied the full notebook and asked for incremental implementation steps and checks. The model helped identify where to start in an unfamiliar codebase, but its proposed code had bugs; the speakers said its main benefit was reducing orientation effort, not necessarily saving time, and emphasized testing and human review.
SIWC Lesson 6
  • Model releases are becoming more frequent and capable, and agentic systems can perform work rather than just answer questions; the economists caution that benchmark gains matter economically only when they unlock valuable tasks, especially less-specified work that remains harder.
  • As enterprise use of agentic AI grows, token usage is rising while frontier labs compete on price: one speaker cited exponentially increasing token use alongside flat revenue, attributing the gap to price competition and cheaper model offers.
  • The economists remain bullish on AI’s eventual benefits but say extreme OpenAI and Anthropic revenue expectations need productivity gains in the wider economy that they have not yet seen; they expect a more gradual, less extraordinary path.
  • The organizing-intelligence paper discussed in the interview treats organizations, not individual agents, as the relevant economic competitors: current agent swarms mostly parallelize and coordinate poorly, while better coordination could enable different firm structures and change wages and employment. One speaker warns that a coordination breakthrough could cause abrupt labor-market disruption before workers are rematched, while leaving its likelihood uncertain.
  • An interviewer disclosed a “machine age fund” for data centers and other AI infrastructure intended to address supply constraints, while noting that AI is not yet broadly diffused and its company-level productivity impact remains unproven.
What History Tells Us About the AI Revolution | Tyler Cowen & Alex Imas
Jeremy Howard
Profile
  • Jeremy Howard and Jono built a small web app inside Solveit, adding a persistent Python code-execution flow; an LLM-prompt mode was left as a follow-on exercise, not shown as complete.
  • Howard’s engineering thesis is to prototype against real use, then extract reusable abstractions rather than design them upfront; he attributes Answer AI’s reported acceleration without additional hiring to software engineering and better abstractions used with AI.
  • The demo surfaced a security caveat: the presenters recognized that a public route allowing arbitrary code execution on their instance was unsafe, stopped it, and replaced it with a safer setup.
Solveit lesson 4