No Sales? No Product? No Problem. AI Investors Still Eager.

“Ben Spector had an unusual pitch for investors last fall.

 

A Ph.D. student at Stanford University with a highly prized artificial-intelligence background, Spector had no near-term plans to make money and no traditional pitch deck. He didn't even have an idea for a hit AI product.

 

What he did have was a lab, called Flapping Airplanes, a novel idea for training AI models and a zeal to hire talented young researchers eager to tackle AI's biggest problems.

 

Venture-capital firms jumped at the chance to back him.

 

"Small teams of brilliant young people that are able to look at the problem in a new way -- those are the kinds of organizations that actually win," said Spector, 25 years old.

 

Flapping Airplanes -- a reference to the biological cues future AI should take from nature -- is part of a new wave of startups some have dubbed "neolabs," which give priority to long-term research and developing new AI models over immediate profits.

 

The interest in neolabs has skyrocketed as investors hunt for the next OpenAI, which began as a research lab and later became one of the world's most valuable startups.

 

A large subset of top AI researchers believe that models such as ChatGPT and Claude have effectively hit a dead end and will never reach a level of intelligence that matches or exceeds that of humans (top AI companies dispute that view).

 

While there are more than a thousand startups with valuations of $1 billion or higher, the number of neolabs is generally believed to be in the dozens, according to researchers and investors.

 

Some of the neolabs have seen their valuations soar into the tens of billions of dollars, prompting critics to suggest that most of them will have slim odds of turning a profit or launching a winning product. The labs have created a recruiting frenzy among academics in ways that are drawing promising students away from academia.

 

This month, Flapping Airplanes raised $180 million at a valuation of $1.5 billion from investors including GV, Sequoia Capital, Index Ventures and Menlo Ventures. Ben Spector took leave from his Ph.D. program in September.

 

Humans& in January raised $480 million at a $4.48 billion valuation to build AI systems that help people collaborate. Reflection AI raised $2 billion in October at an $8 billion valuation to build an open-source model. And Periodic Labs, which aims to develop AI to automate scientific research, launched with $300 million in funding in September.

 

There is also Safe Superintelligence, the AI lab founded by Ilya Sutskever, a co-founder and former chief scientist at OpenAI largely credited with inventing ChatGPT.

 

In June 2024, Sutskever said he was starting a new company with one goal: building safe superintelligence. He has raised $3 billion so far, most recently at a $32 billion valuation, and has been unusually direct with investors about his intentions.

 

"The way I would describe it is that there are some ideas that I think are promising and I want to investigate them," Sutskever said on a November episode of the Dwarkesh Podcast, making no promise that such ideas would lead to a breakthrough, a product or revenue. He also said AI is returning to an "age of research" after scaling up from 2020 to 2025.

 

In the past, the most ambitious AI research happened inside academic institutions or corporate research arms like Google's DeepMind. Startups focused on finding applications of that research that could make money.

 

The AI boom has pushed investors toward funding research itself.

 

"A venture-backed lab -- this is a new thing," said Pete Sonsini, co-founder of Laude Ventures. "It's not traditional venture capital."

 

U.S. AI startups raised a record $222 billion last year, according to research firm PitchBook. Investors said they are seeing a rising number of researchers pitching neolabs.

 

Not everyone is convinced these researchers can generate financial returns.

 

"The technical chasm to cross for each of these neolabs is very substantial and I think that risk is very real," said Ashu Garg, a general partner at Foundation Capital. "The vast majority of them will not cross that at all. They will end up with something that is just incrementally better.

 

And if you're incrementally better than alternatives, you don't matter."

 

One of the biggest challenges neolabs face is talent retention. In an era when CEOs of the largest tech companies are offering more than $300 million to hire AI experts, it is difficult for startups to hold on to their prized researchers.

 

It is a reality Thinking Machines Lab recently highlighted in dramatic fashion.

 

Co-founded by former OpenAI executive Mira Murati, Thinking Machines lost two of its founders, Barret Zoph and Luke Metz, to OpenAI in January. In October, another one of its founders, Andrew Tulloch, departed for Meta. Thinking Machines has sought additional capital in recent months that could value the company at $50 billion.

 

These losses have rattled investors, who have sought to be more probing with AI founders about their incentives.

 

"Is it a financial motivation or is it a motivation to really drive an impact?" said Dave Munichiello, a managing partner at GV and an investor in Flapping Airplanes. "Are they in it for 10 years? Or do they have four houses that they need to pay?"

 

Flapping Airplanes' strategy to compete in the brutal talent war is to not try to hire the most famous researchers. Instead, they are recruiting newcomers who would have ordinarily pursued a Ph.D. or a role at a quant firm.

 

The rush of younger AI startup talent means fewer purely academic researchers. Stefano Ermon, a computer-science professor at Stanford University, said this is the most turnover he has seen in academia in the decade he has been teaching.

 

"There will be fewer people going into academic positions and maybe it will be harder to train the next generation," Ermon said.

 

At the same time, researchers are realizing that the opportunity to raise big VC dollars quickly and easily might not last long.

 

In November, Ermon announced he raised $50 million for a neolab called Inception focused on developing diffusion models [1] to generate text and code.

 

"This is the first time I felt like, yeah, the upside is so big and we are so uniquely positioned to go after this," Ermon said. "It's now or never."” [2]

 

Sutskever, by not promising that his ideas will lead to a breakthrough, product or money, is doing the right thing. Theranos’s advanced unsuccessful concept of using a single drop of blood for a comprehensive diagnosis led to Theranos executives imprisoned for misleading investors. Theranos promised great value in the form of easy diagnostics. Research projects seek to find out what is not yet known, and often, by the definition of research projects, do not yield the expected result. Prison is not the most fun place in the world. 

 

1. Diffusion models are powerful AI generative models that create high-quality data (like images, audio, text) by learning to reverse a gradual "noising" process, effectively turning pure noise into coherent, realistic content through iterative denoising steps, underpinning famous tools like DALL-E and Midjourney by mastering complex data distributions. They work in two phases: a forward process that adds noise to data until it's pure static, and a reverse process, where a neural network learns to predict and remove that noise step-by-step, generating new samples from random noise.

 

How They Work (The Two Processes)

 

    Forward Diffusion (Noising):

        Takes real data (e.g., an image) and gradually adds small amounts of random noise (like Gaussian noise) over many steps.

        Eventually, the original data is completely destroyed, becoming indistinguishable from pure random noise.

    Reverse Diffusion (Denoising/Generation):

        A neural network (often a U-Net architecture [3]) is trained to reverse this process, learning to predict and remove the noise added at each step.

        To generate something new, the model starts with random noise and iteratively applies the learned denoising steps, guided by the network, until a clear, new data sample emerges.

 

Key Characteristics & Advantages

 

    High-Quality Generation: Produces outputs with impressive detail and realism, often surpassing older models like GANs.

    Stability: Offers more stable training compared to some other generative models.

    Versatility: Applicable to various data types, including images, audio, and text.

    Conditional Generation: Can generate specific outputs based on text prompts (text-to-image) or other conditions.

 

Examples & Applications

 

    Text-to-Image: Stable Diffusion, DALL-E 2, Midjourney, Google Imagen.

    Image Editing: Inpainting (filling missing parts), outpainting (extending images).

    Scientific Research: Material science (predicting crystal structures), Earth system science (satellite data).

 

2. No Sales? No Product? No Problem. AI Investors Still Eager. Clark, Kate.  Wall Street Journal, Eastern edition; New York, N.Y.. 28 Jan 2026: A1. 

 

3.  U-Net is a U-shaped convolutional neural network (CNN) architecture, famous for precise image segmentation, especially in medicine, that efficiently captures context and localization by using a contracting (encoder) path to extract features and a symmetric expanding (decoder) path to reconstruct the segmentation map, connected by "skip connections" that fuse fine-grained detail with broader context. This design allows it to work well with limited training data, producing pixel-level classifications (e.g., separating tumors from healthy tissue). 

Key Components

 

    Encoder (Contracting Path):

    Downsamples the input image through convolutional and pooling layers, capturing high-level features (what's in the image).

    Decoder (Expanding Path):

    Upsamples the feature maps, using transposed convolutions [4], and combines them with features from the encoder via skip connections to precisely locate objects (where things are).

    Skip Connections:

    Concatenate feature maps from the encoder directly to the corresponding layers in the decoder, preserving spatial details lost during downsampling, crucial for accurate boundaries.

    U-Shape:

    The symmetrical encoder-decoder structure gives the network its "U" shape, allowing it to learn both context and precise localization.

 

How it Works (Simplified)

 

    Input: An image (e.g., a medical scan) is fed in.

    Encoding: The image goes through stages of feature extraction, getting smaller in size but richer in semantic information (like "this area is likely a cell").

    Bottleneck (Bridge): The deepest part of the U connects the encoder and decoder.

    Decoding: Features are upsampled, and crucial details from the encoder's early layers (copied via skip connections) are added back in to help pinpoint boundaries.

    Output: A segmentation map where each pixel is classified (e.g., foreground/background, tumor/non-tumor).

 

Applications

 

    Biomedical Imaging: Segmenting cells, tumors, organs in X-rays, MRIs, microscopy.

    Satellite Imagery: Mapping roads, buildings, glaciers.

    Generative AI: Image denoising, image-to-image translation (e.g., turning sketches into photos).

 

4. Transposed convolution (also known as deconvolution or fractionally-strided convolution) is a neural network operation used to increase the spatial resolution (upsample) of feature maps, acting as the reverse of a standard convolution. Instead of mapping many input pixels to one output pixel, it maps one input pixel to many output pixels, making it ideal for decoder networks, image segmentation, and Generative Adversarial Networks (GANs). It works by multiplying each input value by a trainable kernel, placing these resulting matrices onto a larger output grid, and summing the overlapping regions. 

Key Aspects of Transposed Convolution: 

• Purpose: To upscale or "decompress" a smaller, low-resolution feature map into a larger, high-resolution one.
• Mechanism: Each element in the input tensor is multiplied by a learnable filter (kernel), and the resulting weighted kernels are placed onto an output map, typically with overlap.
• Stride Interaction: A stride >1 in a transposed convolution does not mean skipping input pixels; rather, it means moving the kernel faster across the output, resulting in a larger output stride.
• Components: Often involves padding, which can shrink the output size, opposite to standard convolution behavior.
• Use Cases: Essential in architectures needing upsampling, such as U-Net for semantic segmentation or GANs for generating images from noise vectors. 

Differences from Regular Convolution:
While a normal convolution reduces a 4×4 input to a 2×2

output, a transposed convolution reverses this, turning a 2×2 into a 4×4 (assuming a 3×3

kernel and a stride of 2). It is important to note that a transposed convolution is not a direct mathematical inversion of a regular convolution, but rather a way to achieve an "upsampled" version of the same shape.