Much of AI interpretability work begins after a model has already been trained. Guide Labs is taking the opposite approach: building interpretability into the model from the start.

In this episode of High Bit, Brett Gibson talks with Julius Adebayo, cofounder and CEO of Guide Labs, about why post-doc tools struggle to explain models with billions of parameters, and how changing the training process can make AI systems easier to understand, audit, and control without the performance trade-off the field has long assumed.

Julius explains how his research challenged the long-held assumption that interpretability and model performance are fundamentally at odds. He also gets into Guide Labs’ move from next-token prediction to diffusion language models, why that architecture may offer greater control, and what breaks when you scale interpretability from hundreds of concepts to tens of thousands.

The conversation covers the unexpected engineering challenges behind training interpretable models, from preventing interpretability loss functions from destabilizing training to dealing with duplicated training data, and Guide Labs’ goal of making every output traceable to its context, the concepts influencing it, and the training data behind it.

Chapters:

(00:00) Change the way you train your model

(00:59) What Guide Labs builds

(01:56) Why interpretability has been hard

(05:33) Julius's background: loans, X-rays, and a PhD

(07:04) The trade-off the field assumed was real

(08:01) Why he stopped trusting post-doc tools

(13:51) The three things that made them commit

(18:01) YC, the first models, and a team of nine

(20:19) Why language models were the hardest test

(23:10) Leaving next token prediction for diffusion

(26:41) Why GPT-3 wasn't just scaling it up

(28:00) 200 concepts by hand, 50,000 you can't

(31:25) Training lore and "divine benevolence"

(32:33) Setting a minimum bar to keep shipping

(34:33) How open to be about your own techniques

(35:51) The standard dataset full of duplicates

(39:31) Where AI coding tools stop helping

(41:43) What's next: the model they're training now

Follow Julius and Guide Labs:

X: @juliusadml / @guidelabsai

LinkedIn:

Julius https://www.linkedin.com/in/juliusadebayo/
Guide Labs https://www.linkedin.com/company/guide-labs/

Follow Brett and Initialized:

X: @brettdg / @Initialized

LinkedIn:
https://www.linkedin.com/in/brettdgibson/
https://www.linkedin.com/company/initialized-capital/

For the first-ever live recording of High Bit, Brett Gibson sits down with two founders building at the center of the AI engineering shift: Daksh Gupta, co-founder and CEO of Greptile, and Matt Matheson, co-founder and CTO of Trunk.io.

AI can write code. That part is no longer interesting. What matters now is what happens when engineering teams suddenly ship much faster, when review and testing become the bottleneck, and when the old operating model for software teams starts to break.

In this episode, Daksh and Matt get into how their teams are structured, why smaller teams can now do much more, what happens to project management when writing a ticket can take as long as shipping the feature, and why the best engineers in this era may look different than before.

They also discuss the tooling layer beyond code generation, from AI code review and CI to agents, testing, validation, and the “outer loop” required to get code safely into production.

The conversation gets into what engineering leaders should rethink now: team size, autonomy, hiring, process, token spend, and how to build teams that can move at AI speed without turning the product into a mess.

Chapters:

(00:00) Advice for Engineering Leaders
(01:07) Welcome to High Bit Live
(02:20) Matt's Path Through Microsoft, Uber, and Trunk
(03:08) Daksh's First Job Is Running Greptile
(04:00) Why Three-Person Teams Beat Everything Else
(07:55) Why AI-Native New Grads Changed Their Minds on Junior Hires
(11:41) When Tickets Take Longer Than Shipping
(16:05) Review and Testing Became the New Bottleneck
(19:40) Greptile's Three Questions for Code Review
(22:34) The Outer Loop Is the Next Bottleneck to Fix
(25:33) The Rise of the God Agent
(29:53) One Tool Led to 40 Internal Workflows
(36:05) The 10,000x Engineer and the War for Talent
(39:10) Two Hiring Archetypes That Beat Generalists
(40:16) Why You Can't Make a Heavy Company Light
(44:40) Why They're Not Maxing Out Token Spend
(49:07) Why Small Inefficiencies Now Cost 10X More

Follow Daksh and Greptile:

@dakshgup

@greptile

Follow Matt and Trunk.io:

linkedin.com/in/matthewmatheson

@trunkio

Follow Brett and Initialized:

X: @brettdg / @Initialized

LinkedIn:

linkedin.com/in/brettdgibson

linkedin.com/company/initialized-capital

Integration is a dirty word in defense. The consultancy model that solved it for decades no longer works when you're dealing with hundreds of systems from dozens of vendors across land, sea, and airspace, all needing to work together within milliseconds.

In this episode of High Bit, Brett Gibson talks with Zane Mountcastle, cofounder and CEO of Picogrid, about how they built the hardware and software infrastructure layer that makes hundreds of defense systems work together.

Zane gets into what it actually looks like to deploy in the field, from detecting a quadcopter 50 feet off the ground, operating in GPS-denied environments with decades-old hardware and adversaries trying to trick your sensors.

Speed became their biggest differentiator. Their average integration time is measured in hours, and what used to take six to twelve months gets done over a weekend.

He also explains why trust is the biggest moat in defense, and how AI is now core to how they build, including training a model on every system they've ever connected so a first pass integration happens in seconds.

Chapters:

(00:00) "Integration is a dirty word"

(00:19) What Picogrid builds

(01:08) How Zane ended up working with the Pentagon

(03:11) The Last Supper: how defense consolidated

(05:01) Why hundreds of new defense vendors made integration worse

(07:22) Why the consultancy model no longer works

(09:41) Why speed is their biggest differentiator

(11:53) Each system has a language

(15:49) Four types of location data to find and track a drone

(19:54) Why Picogrid will never build a drone or sensor

(22:16) The military loves to buy widgets

(30:38) Trust is the biggest moat in defense

(34:20) Calibrating in the field: the camera that finds the sun

(37:36) First team in, last team out

(45:37) How AI is now part of everything they build

(47:00) $45M from Bessemer and what comes next

Follow Zane and Picogrid:

X: @zanemountcastle / @Picogrid

LinkedIn:

linkedin.com/in/zanemountcastle

linkedin.com/company/picogrid

Follow Brett and Initialized:

X: @brettdg / @Initialized

LinkedIn:

linkedin.com/in/brettdgibson

linkedin.com/company/initialized-capital

The internet was built without identity at the protocol level. Now AI makes it nearly impossible to know who or what you're interacting with online. Alien is a decentralized unique identity platform built on the principle of one human, one account — no government required, no central authority, just a cryptographically verifiable way to prove you're human and link that identity to the AI agents acting on your behalf. Mainnet is live now.

In this episode of High Bit, Brett Gibson talks with Kirill Avery, founder and CEO of Alien, about why identity has always been the missing layer of the internet, how Alien combines biometrics, social graphs, and verifiable credentials into a single trust framework, and why the company built its own blockchain rather than deploy on Ethereum or Solana. Kirill also explains why 50 countries already have CBDCs and why building an alternative is the only real answer, what it looks like when one agent manages a thousand agents each managing a thousand more, and why proving you're human online will always be probabilistic — never absolute.

Chapters:

(00:00) Are you human? Why the answer is always probabilistic

(00:19) What Alien is and why unique identity matters

(01:08) Kirill's background: VK, bots, and the identity problem

(03:11) Why the internet was built without an identity layer

(05:01) How AI made the identity problem urgent

(06:24) Why Alien needed its own blockchain

(07:22) CBDCs, centralized identity, and why an alternative is necessary

(10:30) Consensus mechanisms as voting machines and Bitcoin's honesty problem

(12:15) How Alien ensures one person, one account

(16:04) Biometrics, birthdates, and social graphs as layered verification

(19:54) The Alien coin and how value accrues to participants

(22:39) Why early adopter concentration is a design flaw

(27:42) Agent hierarchies and why human-linked agent identity matters

(30:06) How Alien uses AI internally to build with a smaller team

(35:15) What's next: Mainnet phases, Alien coin, and Solana integration

Follow Kirill and Alien for more:

X:@kirillzzy@alienorg

LinkedIn:

linkedin.com/in/kirillzzy

linkedin.com/company/alienorg

Follow Brett and Initialized:

X: @brettdg / @Initialized

LinkedIn:linkedin.com/in/brettdgibson

linkedin.com/company/initialized-capital

Azura built a unified application for trading across multiple blockchains from one interface. You can trade tokens across Solana, Ethereum, Base, Arbitrum, Blast, and more without jumping between wallets, bridges, and apps. Try it: app.azura.xyz

In this episode of High Bit, Initialized managing partner Brett Gibson talks with Jackson Denka, founder and CEO of Azura, about building an onchain trading platform that aims to feel more like a traditional brokerage while staying self-custodial. They dig into why crypto trading is still fragmented — and what Azura is doing about it.

Jackson explains why DeFi is powerful but still hard to use, what it takes to unify DEXs and bridges into a single trading experience, and why cross-chain execution is such a difficult technical problem. They also discuss why he believes value in crypto will accrue at the application layer, why Azura built more of its stack in-house, and what it would take for crypto to disappear into the backend.

Chapters:

(00:00) One venue for every asset

(00:46) What Azura builds

(01:07) Jackson’s origin story

(03:13) Why DeFi is hard to use

(05:39) Why the app layer wins

(09:23) Why onchain is next

(13:05) One app across chains

(15:45) Why this is so hard

(17:12) Cross-chain execution

(21:12) Hiring in DeFi

(22:36) App revenue and DEX growth

(25:44) Removing gas and networks

(27:01) Self-custody, simpler UX

(30:07) Building the stack in-house

(33:13) How Azura uses LLMs

(39:52) What’s next for Azura

Follow Jackson and Azura:

X

@jacksondenka

@AzuraTrade

LinkedIn

https://www.linkedin.com/in/jacksondenka/

https://www.linkedin.com/company/azuraxyz/

Code can be generated faster than ever. But getting that code to actually look good is a different problem entirely.

In this episode of High Bit, Initialized managing partner Brett Gibson sits down with Daniel Bulhosa Solórzano, cofounder and CTO of Variant, about what it takes to build AI that gets design right, not just code that runs.

Daniel started thinking about this problem in 2017 at Weebly, when the models weren't close to ready. After a stint in self-driving, he came back to it. Variant generates UI that's visually designed, not just technically correct, and shows you multiple options at once so you can find what you actually want instead of having to describe it upfront.

Topics include:

- Why AI struggles with visual quality even when the code works

- The difference between precision and recall in design generation

- Why showing people options beats asking them to describe what they want

- Why people will always want to stay in the loop on design

- How the role of designers changes as AI handles more of the execution

(00:00) What design AI could eventually do

(00:44) What Variant builds

(01:25) How Daniel got here: Weebly, self-driving, and an unsolved problem

(03:20) Why visual code generation is hard

(04:10) Precision vs. recall and why design is different

(06:55) What good design actually means

(09:04) Landing page vs. dashboard: how context shapes design choices

(11:31) What can be described vs. what has to be labeled

(13:10) Casting a wide net for what good design looks like

(15:25) Building the product around an imperfect model

(17:04) Why people react to designs faster than they can describe them

(19:29) Why people won't give up design to AI

(23:40) What AI does to the design learning curve

(25:06) Designers as design managers for agents

(37:42) The 50%/80% horizon and what it means for engineering teams

Follow Daniel and Variant:

Daniel Bulhosa Solórzano

X: https://x.com/bulhosa

LinkedIn: https://www.linkedin.com/in/dbulhosa/

Variant

Website: https://variant.com/

X: https://x.com/variantui

LinkedIn: https://www.linkedin.com/company/variantui/posts/?feedView=all

High Bit

Watch more episodes: https://www.youtube.com/@InitializedCapital

Follow on Spotify: https://open.spotify.com/show/36gTYrH1wlYzTZwLQywbIf

Code can be written faster than ever. But getting that code safely into production is where many engineering teams lose time.

As organizations grow, CI failures, flaky tests, and conflicting pull requests start to compound.

In this episode of High Bit, Initialized managing partner Brett Gibson sits down with Eli Schleifer, founder and CEO of Trunk, to talk about the systems that keep CI green and allow engineering teams to land code reliably as organizations grow.

Before starting Trunk, Eli built developer infrastructure at Microsoft, started a company that was acquired by Google, and later worked with hundreds of engineers at Uber ATG. At Google he saw how powerful internal developer tooling could be. At Uber he saw engineers spend days trying to land code because those systems did not exist. That gap led him to start Trunk.

Eli explains why engineering productivity slows once dozens or hundreds of engineers share the same repository, how flaky tests quietly waste engineering time, and how merge queues prevent broken builds and conflicting pull requests.

Topics include:

• Why CI becomes the bottleneck as engineering teams grow
• How merge queues keep builds reliable
• Why flaky tests waste engineering time
• The build vs buy decision for developer tooling
• How coding tools are increasing pull request volume
• How engineering workflows are changing

(00:00) AI fixing flaky tests

(00:40) What Trunk builds

(02:13) When CI becomes the bottleneck

(02:33) Eli’s background: Microsoft, Google, Uber

(04:29) Why dev tools must solve real pain

(05:09) The merge queue problem

(09:25) Build vs buy for developer tooling

(14:22) How merge queues handle large codebases

(18:58) How coding tools increase PR volume

(21:53) Flaky tests and engineering productivity

(25:24) Using CI and test history to debug failures

(27:37) How engineers prune the search space

(33:23) Engineers as conductors of automated systems

(37:49) What’s next for Trunk

Follow Eli and Trunk:

Eli Schleifer

X: https://x.com/elischleifer

LinkedIn: https://www.linkedin.com/in/elischleifer/

Trunk

Website: [https://trunk.io](https://trunk.io/)

X: https://x.com/trunkio

LinkedIn: https://www.linkedin.com/company/trunk-io/

High Bit

Hosted by Brett Gibson, managing partner, Initialized

https://open.spotify.com/show/36gTYrH1wlYzTZwLQywbIf

AI models keep getting better, but most AI systems still fail in production. Why?

In this episode of High Bit, Brett Gibson sits down with Ghita Houir Alami, cofounder and CEO of ZeroEntropy, to break down the real bottleneck holding AI agents back: retrieval.

Ghita explains why embeddings alone can’t reliably surface the right information, why tools like Slack search feel so frustrating, and how rerankers add a critical second pass that dramatically improves accuracy. She walks through ZeroEntropy’s approach to training rerankers using pairwise comparisons and Elo-style scoring, and why this method generalizes across domains like code, finance, and biology.

The conversation goes deep into:

• Why AI agents fail even when the data exists.
• How reranking fixes poor ordering from vector search.
• Why “accuracy” now includes helpful context, not just correct answers.
• What actually changes when retrieval becomes trustworthy enough to remove humans from the loop.

If you’re building AI agents, search systems, customer support bots, or internal knowledge tools, this episode explains what’s breaking today, and what has to change for AI to work reliably at scale.

(00:00) What changes when retrieval works

(00:39) What ZeroEntropy builds

(01:42) Why retrieval became the real problem

(03:12) Why search fails (Slack included)

(05:11) Why embeddings fall short

(07:11) Rerankers: the missing layer

(10:11) Why rerankers matter most

(12:44) Pairwise ranking vs scoring

(13:52) Elo scoring for documents

(16:33) Fast rerankers via distillation

(18:07) Why old training methods break

(21:29) Retrieval for AI agents

(24:20) Recency, memory, personalization

(32:06) What reliable retrieval unlocks

(33:42) What’s next for ZeroEntropy

Follow Ghita and ZeroEntropy for more:X@ghita__ha@ZeroEntropy_AILinkedInhttps://www.linkedin.com/in/ghita-houir-alami/https://www.linkedin.com/company/zeroentropy-inc

Electrification isn’t easy—and most people don’t see the chaos beneath the solar panels, batteries, EV chargers, and heat pumps going onto the grid. Coperniq cofounder and CTO Max Kazakov breaks down the hidden workflows behind distributed energy: legacy tools installers still rely on, hardware that doesn’t want to integrate, and why the next generation of “utilities” will look nothing like the last.

Coperniq is the workflow platform for contractors and energy companies to move from post-its and spreadsheets to a system that sells, permits, installs, and maintains distributed energy assets over decades.

Max also shares what it takes to build vertical SaaS for the physical world: curbside Figma demos during COVID, rebuilding their mobile app for 120°F rooftops with no cell service, designing a workflow engine that matches real-world permitting and interconnection, integrating a wild west of OEM hardware, and how AI is already reshaping their product and engineering culture.

Content:

(00:00) The Invisible Glue of the New Grid

(01:05) The Second Electrification Wave

(02:51) Cofounder Origins: Russia, Yemen, Berkeley

(06:12) Humans + Hardware Coordination Challenge

(08:10) Anti-MVP: Mini ERP on Day One

(11:56) Curbside Figma Demos during COVID

(14:47) Field Reality: 120° Rooftops, Zero Cell Service

(20:03) Stateful Workflows (Permits, Interconnection, Construction)

(24:56) Integrating OEM Hardware (Hitting Walls)

(28:41) The Dongle Question: Do we need software afterall?

(31:25) Rebuilding the Mobile App for an Offline-First World

(36:40) Hire Tinkerers, Not Pedigrees

(43:00) How AI Is Reshaping Coperniq

Subscribe to High Bit for more conversations with technical founders building the future.

Follow @Coperniq_AI for more.

Robots built like humans.

On High Bit, Dhanush Radhakrishnan, cofounder & CEO of Clone, explains how they’re letting biology set the blueprint for musculoskeletal, super-intelligent androids — synthetic humans straight out of sci-fi.

Powered by artificial muscles instead of motors and attached to anatomically accurate skeletons, Clone is building robots designed for human-level motion, durability, and full-body control.

Dhanush explains the early engineering choices that helped them move fast, their data strategy (motion capture, teleoperation, egocentric video), and his excitement about a future untethered biped.

Hosted by Brett Gibson, managing partner at Initialized.

Content:

(00:00) From Fragile to Durable

(00:37) Musculoskeletal Androids

(02:03) Why Artificial Muscles

(04:10) Starting Clone in Poland

(06:29) Removing Early Sensors

(10:14) The Durability Challenge

(14:10) Anatomy as Blueprint

(16:22) Building Custom Valves

(18:17) Hand to Full Body

(24:40) Prototyping with Pneumatics

(29:26) Delaying Tactile Skin

(32:39) Data: MOCAP + Teleop

(45:20) Toward an Untethered Biped

Subscribe to High Bit for more conversations with technical founders building the future.