STARTUPS

Pathway just challenged AI’s scaling economics

One of the first "neolabs" to announce something tangible is showing off an AI breakthrough that would fundamentally change the architecture of today's AI, making it cheaper to operate and requiring far less data center power. 

Pathway unveiled a 150-million parameter small reasoning model, BDH-CQ, on Tuesday, along with benchmarking results that back up Pathway's claims that its post-transformer architecture could deliver comparable performance at a fraction of the cost and computing resources of today's leading frontier models. 

According to the ARC-AGI-1 benchmark, BDH-CQ achieved 29.5% pass@2 accuracy (it solved nearly a third of the problems on the test when given two guesses) with a computed inference cost of $0.0007 per task. So how does that compare with OpenAI's most cost-effective model? GPT 5.6 Luna (Low), which OpenAI just reduced in price by 80% on July 30, scored 34.5% on the same benchmark. However, even at its new cut-rate price, it cost 11 times more than Pathway's new model. 

Part of that is because the Pathway model is small, doesn't need chain-of-thought to achieve reasoning, and needs less data because of its improved memory. So Luna has slightly better performance at an astronomically more expensive price. And Luna itself is a fraction of the price of the leading frontier models. So while it's very still early, what Pathway has achieved holds tremendous promise for future efficiency and cost reductions of frontier-class models.

"We need to be able to squeeze more intelligence per dollar, and for this you need to change the paradigm," Zuzanna Stamirowska, CEO and co-founder of Pathway, told The Deep View. "This is a very deep innovation, and we wouldn't have done it if it wasn't going to be, and if it didn't have a chance to capture the market."

The Pathway team believes their breakthrough is "a PageRank moment for intelligence," referring to the turning point when Larry Page and Sergey Brin realized they could make web search dramatically better by ranking pages based on the structure of links between them and not just the keywords on the page.

Stamirowska, who also appeared on The Deep View Conversations this week to do a deep dive on the fundamental limitations that are holding back LLMs, has used her background in research and game theory to assemble a team of researchers and advisors with impressive achievements:

• Łukasz Kaiser: co-author of the original paper on the Transformer that launched the generative AI revolution; serves as an advisor to the company, also independently verified the ARC-AGI-1 benchmark

• Alex Kurzok: former group product manager of Gemini at Google DeepMind, now chief product officer at Pathway

• Jonathan Frankle: chief AI scientist at Databricks (who spoke with The Deep View recently) serves as an advisor to Pathway on scaling and deployment and is also an investor in the company

• Martín Farach-Colton: chair of computer science and engineering at NYU and one of the early employees of Google who led several important technological breakthroughs; now serves as an advisor to Pathway on its scientific vision

In October 2025, the Pathway team first published its paper, "The Dragon Hatchling: The Missing Link between the Transformer and Models of the Brain," which quickly gained an audience in the AI research community. It laid out the company's theory and vision for a post-LLM future (BDH stands for "beautiful dragon hatchling" and is a sci-fi/fantasy reference from a Terry Pratchett story). The new benchmark released around the BDH-CQ model is the first evidence that Pathway's theory is on the mark. The team at Pathway still has a lot of work to do before its models can compete with the leading edge frontier models, but they are convinced their architecture can scale to 600B parameter models that will compete with the world's leading models and will change the economics of AI. The scaling laws that power today's top models consume so much compute, energy, and data and they are hitting limits. The industry desperately needs something like breakthroughs Pathway is espousing. There are over 40 neolabs that have raised $40B in funding and they are attacking these problems. Whether or not Pathway is the eventual flagbearer, keep an eye on these startups as major disruptors in the months and years ahead.

TOGETHER WITH AWS RE:INVENT

AWS re:Invent 2026 is designed for hands-on problem-solving

Over 2,200 sessions, 70% are interactive. re:Invent is returning to Las Vegas, November 30 - December 4.

You’ll have the opportunity to: 

• Run a service against a realistic workload before it hits production

• Ask a service engineer about the constraint that's actually blocking you

• Access AWS certification discounts, with exam prep built into the schedule

• Talk through implementations with 400+ partners on the expo floor

Learn more and sign up by August 25th to save $1,200.

BIG TECH

Why Meta is launching back into open models

As the momentum around open models continues to rise, Meta has unexpectedly re-entered the mix. 

On Monday, the company unveiled Muse Glimmer, an open weight, agentic model that's small enough to run on user devices. The 30-billion parameter model can operate on a Mac or PC with a "single consumer GPU," Meta said, and is available under the permissive Apache 2.0 license. 

Meta said that the model offers strong performance in agentic use cases compared to leading models of similar sizes, outperforming Google's Gemma4-31B and Qwen3.6-27B on benchmarks for things like coding, multi-step tasks, and software engineering. 

Some of Muse Glimmer's features include: 

• End-to-end agentic task completion

• Reliable tool use

• Multi-step reasoning

• And multilingual and multimodal input and reasoning

In its blog post, Meta said that small, local models will enable users to "use AI anywhere, anytime, with or without an internet connection," rather than relying on cloud infrastructure and network access. Additionally, Muse Glimmer being an open model means that the US "finally has its response to the open weights AI race," Aaron Levie, CEO of Box, said in a post on X. "This will continue to help drive down the cost of intelligence." 

That's a hyperbole since Nvidia, OpenAI, and Google all have respectable open models and Mira Murati's startup Thinking Machines has recently released its open model. 

The release of Muse Glimmer falls directly in line with Meta's broader pitch for democratizing personal superintelligence, a topic that CEO Mark Zuckerberg went into detail on in a 6,500-word blog post on Monday. 

The post, titled "The Future is for Everyone" and similar to past manifestos from Anthropic's Dario Amodei and OpenAI's Sam Altman, raises concerns about the concentration of power in AI as we race towards superintelligence, claiming that the notion that AI should be "centralized and restricted to a few institutions" is "inherently problematic." 

"We propose a philosophy based on individual empowerment as the source of prosperity, invention as the primary purpose of superintelligence, and balance of power as the foundation of safety," the billionaire wrote.  

In this essay, Zuckerberg makes promises about how Meta is approaching superintelligence, including giving everyone an "exceptionally capable personal agent" that knows everything about you, "tools for creation" to express ideas, create new businesses or contribute to scientific progress, personalized tutors and coaches, and free and affordable access to these tools. Zuckerberg also circled back to open-source AI, noting that the US must rethink its open-source policies to enrich the ecosystem.

Muse Glimmer is a welcome addition to the market. It gives users the ability to run AI locally on-device without needing to rely heavily on outside compute power or proprietary APIs. It contributes to the open-source ecosystem in the United States, allowing users to have more options to choose from that aren't just proprietary frontier models or Chinese open weight competitors. And it may help Meta get back on its feet as it lags behind AI rivals. Still, Zuckerberg proselytizing about the dangers of centralization in AI and his goal for proliferating personal intelligence might set off some alarm bells. It adds to a growing number of prominent voices warning against a concentration of power and influence, while also holding an inordinate amount of power themselves. As AI embeds itself deeper into the fabric of society, the trick is to believe the things that these companies do, not the things that their leaders say. Zuckerberg or Amodei or Altman can all sound the right notes, but their actions are much louder than words.

TOGETHER WITH ATTIO

Attio is the CRM for agentic revenue.

Granola uses Attio for 83% faster lead triage. Modal uses Attio to orchestrate workflow automation at scale. And Railway uses Attio to convert product signals into pipeline.

With Attio, you can run your revenue work exactly as you want. Get agents working on every account, surfacing new opportunities, and handle the work that used to take your team days. 

Whether you're working in your browser, inbox, or favorite agent, connect to your customer data in real-time through Attio's web app, MCP, API, and SDK.

The next era of go to market runs on Attio. Join the 30,000+ teams already there.

Try Attio now, for free.

PRODUCTS

How Pixel photography became an AI proving ground

For over a decade, Google's Pixel phones have used AI and machine learning to leap forward in phone photography. 

The company has pushed the bleeding edge by using software and algorithms to make its phones produce images more and more like a professional camera. While there have occasionally been mixed results, we can't say that the Pixel camera hasn't advanced the boundaries of what's possible. Year after year, the devices have continued to launch new, updated, and refined features to expand and improve phone photography using AI models. 

In an exclusive interview with The Deep View, Isaac Reynolds, who has led the Pixel camera team for the past decade at Google, explained how his team has set its sights on tackling a series of "durable problems" in phone photography:

• Dynamic range

• Low-light performance

• Zoom quality

• Detail and sharpness

• Background blur

• Motion freezing

• Artifacts

• Timing

"We chose durable problems for a reason," said Reynolds, "[because] no matter how good the technology gets, users always want more."

As we prepare for the launch of the Pixel 11, let's do a quick recap of the AI and ML features that have launched on the Pixel phones across the generations: 

• Pixel (2016): Always-on HDR+

• Pixel 2 (2017): Portrait Mode with ML depth estimation in a single lens

• Pixel 3 (2018): Night Sight, Super Res Zoom, Top Shot

• Pixel 4 (2019): Live HDR+, Dual Exposure Controls, computational astrophotography

• Pixel 5 (2020): HDR+ with Bracketing, ML-powered Portrait Light

• Pixel 6 (2021): Real Tone, Face Unblur, Motion Mode, Magic Eraser

• Pixel 7 (2022): Photo Unblur, Guided Frame, improved Super Res Zoom

• Pixel 8 (2023): Best Take, Magic Editor, Audio Magic Eraser, Video Boost

• Pixel 9 (2024): Add Me, Reimagine, Auto Frame, Night Sight Panorama

• Pixel 10 (2025): Camera Coach, Auto Best Take, diffusion-based Pro Res Zoom

One thing Reynolds made clear was that progress has accelerated significantly in the past two years because of generative AI and the integration of Gemini-class models into the team's toolset. 

"Things that we thought were going to be really, really hard are actually really easy," said Reynolds. "I'm seeing examples of models that just work. We are able to scale them much harder, much further, much faster than we ever could have suspected… It breaks our planning process, because things are so much easier now than they were even a year or two ago… I'm looking forward to the next three years. It makes my job a little bit easier when all the things just magically work."

Naturally, one of the biggest questions becomes what's actually a photograph? If you can use generative AI to fix the light, make night appear like daytime, add a different sky into a landscape shot, add a person who wasn't in the shot into a group photo, and extend the canvas to add parts to an image that weren't originally there, then a photograph starts to become more like a work of art than a snapshot of reality. Of course, this question goes back to the beginnings of photography itself. Ansel Adams famously talked about how he would manipulate an image in the darkroom so that the photo didn't merely capture what he saw but would portray what it felt like to stand in a particular spot. Today's phones have AI capabilities that give us more and more tools to make art with photography, if we choose to look at it that way. And no phone pushes the boundaries of what's possible more than the Pixel. That makes it exciting to see what comes next, which we'll learn at this week's Made by Google event on August 12. I'll be onsite in New York and will report back the most interesting new developments.

LINKS

• OpenAI expands Project Daybreak with two tiers, GPT‑5.6‑Cyber

• Microsoft reportedly to unveil next-gen AI chip in September

• San Francisco rent is up 18% year-over-year as AI salaries climb

• Ford rolls out AI assistant that checks fuel levels, tire pressure

• Sen. Bernie Sanders calls on frontier AI labs to pause model development

• Applied Compute in talks to raise funding at $3 billion valuation

• Grok Image 2.0: xAI's next-generation image model with improved editing, text rendering and factuality. 

• Claude Sonnet 5: Anthropic has made the introductory pricing permanent at $2 per million input tokens and $10 per million output tokens. 

• Runway: The AI video platform now hosts P-Image-Ideogram, allowing users to go from prompt to image in less than a second. 

• Dyna 2: New from Dyna Robotics, a world-action model pre-trained on one million hours of human video.

• Visa: Analyst, Applied AI Solutions

• Anthropic: Research Engineer, Universes

• Nvidia: ML and Agentic Systems Engineer

• AMD: AI Creative Technologist

GAMES

POLL RESULTS

Do you think large language models will reach a limit in their ability to learn?

Yes (42%)
No (47%)
Other (11%)