Breaking: Guide Labs ships Steerling-8B - the LLM that cites its sources From MIT to YC: Julius Adebayo's interpretability bet Every token, traced. Every output, auditable. Seed round closed. Initialized Capital led. The interpretable era begins. Breaking: Guide Labs ships Steerling-8B - the LLM that cites its sources From MIT to YC: Julius Adebayo's interpretability bet Every token, traced. Every output, auditable. Seed round closed. Initialized Capital led. The interpretable era begins.
Julius Adebayo portrait
Fig. 1 - Subject in his natural habitat: a sentence with a footnote.
Issue No. 037 - The Interpretability Files

Julius
Adebayo

He spent his PhD proving a popular technique for explaining AI didn't actually explain anything. Then he started a company to do it properly.

ROLE Co-Founder & CEO, Guide Labs
BASE San Francisco
LINEAGE MIT / Google Brain / Genentech
Verified
Receipts
Inside
The Profile

A receipt for every word the model says.

Steerling-8B is an open-source large language model with eight billion parameters and an unusual feature: every token it generates can be traced back to the slice of training data that influenced it. There is no current crop of frontier model that does this. Guide Labs, the San Francisco company shipping it, was co-founded in 2023 by Julius Adebayo, a Brigham Young mechanical engineer who detoured through MIT and ended up rebuilding the field he had earlier punched a hole through.

Adebayo's 2018 paper, "Sanity Checks for Saliency Maps," argued that the bright heatmaps researchers used to claim a neural network was "looking at" the right part of an image often failed basic tests. Randomize the model's weights and the heatmap looked the same. A field that thought it could see now had to admit it had been squinting. The paper has been cited thousands of times. It also closed off the easy version of his life's work. If the obvious tools were broken, the only honest path was to build new ones.

The company - and its claim

Guide Labs calls itself the first inherently interpretable AI platform. The phrase is doing work. Most current interpretability tools wrap a model after training - you build the box, then peer in. Guide Labs goes earlier. The models are designed from the start so that their internals correspond to concepts a human can name. The token-level attribution in Steerling-8B is the loud version of this idea. The quiet version is harder to summarize and easier to misstate, so Adebayo, by reputation and habit, tends not to.

We are building interpretable AI systems that humans and domain experts can easily audit, steer, and understand. - Julius Adebayo

Why this, why now

A model that can show its work is not a luxury feature. It is the difference between an AI you can deploy in a hospital and one you cannot, between a regulator clearing a tool and a regulator banning it, between a clinician trusting an output and ignoring it. The current AI moment is loud about capability. It has been quieter about whether anyone other than the model itself knows what is going on inside. Adebayo's bet is that the quiet question will catch up.

His co-founder, Aya Abdelsalam Ismail, is Guide Labs' chief science officer. The two met in the small world of academic interpretability research. The company's small team - sixteen at last count - collects PhDs from MIT, UMD, and Mila, with two dozen papers at top conferences between them. The investor list is the kind that signals "credentialed bet on a hard idea": Initialized Capital led the seed round, with Y Combinator, Tectonic Ventures, Lombardstreet Ventures, Pioneer, and E14 alongside. Reporting from TechCrunch puts the disclosed seed funding around $9.3 million.

8BParameters in Steerling-8B
24+Papers at top venues
2018The "Sanity Checks" year
$9.3MDisclosed seed round

Provo to Cambridge to YC

Adebayo's undergraduate degree, in mechanical engineering, came from Brigham Young University. From there he went to MIT for a master's in computer science and technology policy - an unusually civic pairing for a future AI founder - then stayed for a PhD in computer science under Hal Abelson, the legendary MIT professor who co-wrote the canonical Structure and Interpretation of Computer Programs and co-founded the App Inventor project. Adebayo's doctoral funding came from Open Philanthropy, the same outfit that has bankrolled a large fraction of modern AI-safety research.

Between MIT and Guide Labs sat a year of postdoctoral work at Prescient Design - Genentech's machine learning group - with Kyunghyun Cho and Stephan Ra, applying the interpretability instinct to a domain that genuinely demands it: protein design. His 2025 ICLR paper, "Concept Bottleneck Language Models for Protein Design," is one of several outputs from that stretch. Earlier, before the PhD, he was a Brain Resident at Google and a research engineer at Fast Forward Labs. The CV is a slow accumulation of "people who care about this question."

What the model actually does

Imagine asking an LLM a chemistry question and watching it answer. Now imagine the answer arriving with a faint trail of citations behind every clause - not citations to the open web, but citations to the actual passages in the training data that pulled the answer toward what it became. That is the Steerling-8B demo. Guide Labs' research note "Scaling Interpretable Language Models to 8 Billion Parameters" walks through the architecture. The trick is not bolting attribution on at the end but designing the network so that influence is computable by construction.

The skeptical scientist version of Adebayo will be quick to qualify all of this. Interpretability is not solved. Steerling is not yet a frontier model. Tracing influence is not the same as explaining reasoning. But the field has had a decade of vague claims about explanation. A model with receipts, even imperfect ones, changes the conversation.

Quirks of the record

Adebayo's Twitter handle, @juliusadml, encodes its own thesis - ADML, the initials plus ML. His Google Scholar page reads like a guided tour of the last seven years of interpretability research: saliency map sanity checks, model debugging tools, error discovery by clustering influence embeddings, concept bottleneck generative models, faithfulness-measurable models. Each title is a step in a single argument. The argument is that we should know, with rigor, what models are doing.

For all that, the public version of Julius Adebayo is light on theatre. There is no podcast circuit. The blog is the company blog. The talks are research talks. The vibe, to borrow a current word, is a researcher who became a CEO without losing the researcher.

A receipt, sketched

STEERLING-8B :: OUTPUT TRACE #00237
PromptWhat is X?
Token 1← Doc 4.221
Token 2← Doc 9.018
Token 3← Doc 4.221
Token 4← Doc 1.555
Confidencetraceable

The bigger frame

There is an argument doing the rounds that interpretability is the next scaling law - that after you stack enough parameters and enough data and enough compute, the next axis of progress is whether anyone can understand the result. Adebayo would not say it that boldly. The papers say it for him. So does the company. So, eventually, will the regulators that are circling AI deployment in medicine, finance, and law. If the next decade of AI is a slow handoff from raw capability to trustworthy capability, Guide Labs is camped at the handoff.

For now, the team is sixteen people, the model is open source, the investors are patient, and Julius Adebayo is doing what he has been doing since 2018 - building tools that make a model show its work and then telling you, with rigor, whether the work it is showing is the work it actually did.

Recent receipts

In February 2026, TechCrunch covered the Steerling-8B release as "a new kind of interpretable LLM." Guide Labs' own blog post, "Scaling Interpretable Language Models to 8 Billion Parameters," walks through the architecture choices. Adebayo's MIT thesis, "Towards Effective Tools for Debugging Machine Learning Models," is still hosted on App Inventor's servers, where his adviser Hal Abelson's other students leave their own marks. The thread between thesis and startup is short. The startup is the thesis, with money and a team.

The Scrapbook

Three things that amuse and inform

The Handle

@juliusadml on Twitter/X. ADML = Adebayo + ML. A username that doubles as a research statement.

The Adviser

Hal Abelson co-wrote the book most CS students secretly wish they had finished. Adebayo's thesis lives on the App Inventor server because of him.

The Pivot

Mechanical engineering at Brigham Young to interpretable AI in San Francisco. The shortest path between the two ran through Cambridge.

The Research Arc

An argument, paper by paper.

A small chart of the research line, weighted loosely by where it sits in the through-line from "explanations are broken" to "let's build better ones."

Sanity Checks (2018)
92
Debugging tools (2020)
74
Error discovery (NeurIPS 2023)
68
Concept bottlenecks (ICLR 2024-25)
86
Steerling-8B (2026)
100

A subjective weight, eyeballed from citations + downstream reuse. Useful as a shape, not a number.

"If the obvious tools were broken, the only honest path was to build new ones."
- The shape of his last decade

Share this profile

If you know someone who is building, regulating, or worrying about AI - send them this.

LinkedIn Twitter / X Facebook Instagram