James Bowler, Head of Research Partnerships at AE Studio, is joined by two guests to explore emergent communication in multi-agent AI systems: what happens when agents under pressure develop communication protocols no human designed and few humans can read. Hale Sirin leads the AI agents program at Schmidt Sciences' AI and Advanced Computing Institute and is an assistant research professor at Johns Hopkins. Elias Stengel-Eskin is an assistant professor at UT Austin and a lead PI on the program.

The core concern is under-appreciated: as more agents are deployed by more actors across shared infrastructure, how those agents talk to each other matters as much as what any individual agent can do. Hale frames this through Schmidt Sciences' agents program, a pilot studying how inter-agent communication evolves, stabilizes, or diverges. The worry isn't just that agents become more capable, but that their protocols may become opaque, making human oversight impossible exactly when it's most needed. Elias adds a second motivation: these systems give linguistics a new subject of study, a chance to watch meaning and even syntax change through interaction.

To study this rigorously, the team, led by Elias and language evolution expert Simon Kirby (University of Edinburgh), built a test bed around a deliberately out-of-distribution scenario: a sci-fi medical emergency involving a fictional alien patient called Veyru. Because LLMs have no training data on Veyru's anatomy, the specialist genuinely knows things the field agent does not, forcing real communication rather than single-agent collapse. Under token-budget pressure, agents consistently developed inscrutable protocols, not just abbreviation dictionaries but compositional languages where message order encodes meaning, the kind a cryptanalyst would need grounding data to decode. The key ingredient was a postmortem stage, an unconstrained channel after each round where agents could deliberate on what went wrong. No one told them to build a language there. They did it anyway.

Two results stand out. First, the same pair of models produces meaningfully different languages across runs, providing the variability needed to study transmission: which protocols survive when a new agent is swapped in mid-task, like a shift change in an emergency room. Second, because these are LLM-based agents, newcomers can take an active role in language acquisition, asking clarifying questions when a symbol is too ambiguous to infer and updating their model before acting. This active learning dynamic is largely absent from earlier literature. The conversation closes on how AE Studio's research engineers, paired with academic leads, can accelerate research and build open source infrastructure for the wider community.

In this episode:

• What emergent communication is and why it matters for AI oversight
• How information asymmetry between agents drives genuine communication
• What emergent communication reveals about how LLMs understand language
• How a postmortem channel pushes agents from abbreviations to compositional protocols
• Why compositional languages are harder to decode but easier to transmit to new agents
• How LLM agents actively probe for missing vocabulary
• The Schmidt Sciences open call, Scaling AI Safety for a Multi-Agent World
• Why multi-agent setups are necessary for partial goal alignment or asynchronous execution

Learn more: https://ae.studio/alignment

Read the blogpost: https://www.schmidtsciences.org/glossogen/

AE Studio is hiring: https://www.ae.studio/join-us

Subscribe to our newsletter: https://aestudio.beehiiv.com/

Schmidt Sciences: https://www.schmidtsciences.org/

James Bowler LinkedIn: https://www.linkedin.com/in/james-bowler-84b02a100/

Hale Sirin LinkedIn: https://www.linkedin.com/in/hale-sirin/

Elias Stengel-Eskin LinkedIn: https://www.linkedin.com/in/elias-stengel-eskin/

Explore GlossoGen: emergentcomms.ai

Contact us: [email protected]

In this episode, James is joined by Dan Balsam, co-founder and CTO of Goodfire, to talk about Silico, Goodfire's new automated AI research platform, and why the alignment problem is too big to leave to the frontier labs alone.

They start with what interpretability has already delivered: by reverse-engineering what a genomics foundation model represents internally, Goodfire and Mayo Clinic achieved state-of-the-art variant effect prediction on a model that had never seen human variation at all. These results have since been confirmed experimentally by external labs. Dan explains why predicting the next nucleotide across every sequenced organism forces a model to learn the structure of evolution itself, and why most of a model's capabilities are already there after pre-training.

From there they take on the so-called forbidden techniques. Goodfire's reward shaping research uses a signal from a frozen copy of the model inside the reward function, and it has been scaled past the trillion-parameter point on open models like Kimi. Dan walks through the standard objections, that applying optimization pressure against an interpretability probe just teaches the model to hide its scheming rather than stop it, and why he thinks simplicity bias and shared circuitry make upstream intervention more robust than the philosophical argument assumes. His position is that whatever solves alignment will involve grabbing the steering wheel of the training process, and that no genre of technique should be off-limits before we have the empirical evidence.

The back half is a deep dive on Silico: a multi-agent harness attached to real compute that runs long-horizon research autonomously over days or weeks, and the product work behind closing the trust gap when users let agents run unattended. AE Studio's research team uses it daily. Dan and James close on where Goodfire sits in the ecosystem. They cover why self-auditing labs are untenable in the long run, the economic case that alignment pays for itself, and Dan's call that we are in something like wartime for alignment.

In this episode:

* Why Goodfire says no alignment technique should be "forbidden"

* How reward shaping works, and how it scales past trillion-parameter models

* What Silico is and how it runs long-horizon research on its own

* How decoding a genomics model produced state-of-the-art disease-variant prediction

* Why the big labs can't solve alignment alone

* The economic case that safety and alignment pay for themselves

* Silico grants for AI safety researchers and academics

Learn more: ae.studio/alignment

Silico: https://goodfire.com/silico

AE Studio is hiring: https://www.ae.studio/join-us

Subscribe to our newsletter: https://aestudio.beehiiv.com/

Dan Balsam LinkedIn: https://www.linkedin.com/in/dbalsam/

James Bowler LinkedIn: https://www.linkedin.com/in/james-bowler-84b02a100/

Contact us: [email protected]

In this episode, James Bowler is joined by Melanie Plaza, Chief Technology Officer at AE Studio, to explore the commercial case for AI alignment, arguing that the techniques researchers care about most are the same ones that make production AI systems actually trustworthy and valuable.

Melanie draws on a decade of applied AI delivery to show where alignment and commercial engineering have converged. Getting an agent to do a thing is trivial now; getting it to do the right thing consistently enough to trust in production is not. The gap between those two states is filled by the same tools alignment researchers reach for: rigorous eval suites, red teaming, carefully specified behavior, and guardrails that hold up under adversarial pressure. Melanie's observation is that teams who skip this work don't just expose themselves to safety risk, they fail to get ROI, accumulating what she calls "a sprawl of pilot death."

The conversation gets concrete with a case AE Studio has been working on: AI characters, some of them villains, interacting directly with users including children in open-ended multi-turn conversations. Moving from tightly scaffolded, stepwise pipelines to goal-oriented prompting produces better, more natural results, but it also opens a much larger space of possible outputs. The team had to develop a taxonomy of harm categories, including direct user harm, brand harm, and something they call character taboos, illustrated by the example of Peppa Pig cheerfully recommending bacon. That particular output won't end the world, but it points at a real problem: the tail of a generative system is enormous, and standard single-turn test suites won't find what lives out there.

Melanie and James also work through the economics of open-weight models, noting that as models like GLM 5.2 and Kimi K3 close the capability gap, the cost and dependency arguments for closed-source APIs weaken. That opens the door to fine-tuning, RL, and white-box techniques like activation steering, which could address problems that prompt-level guardrails handle poorly. It also shifts safety responsibility onto the deploying organization, since the model-level protections that frontier labs build in by default no longer come for free. The upside is high; so is the risk if teams treat alignment as an afterthought.

In this episode:

* Why the hardest problems in applied AI deployment are alignment problems under a different name

* How the shift from scaffolded pipelines to goal-oriented, multi-agent systems removed a de facto safety constraint that few teams replaced

* The harm taxonomy AE Studio built for AI characters interacting with children, including the "character taboos" category

* Why standard single-turn eval suites miss the long tail of multi-turn conversations

* The economic case for open-weight models and what safety responsibilities transfer to the deploying organization when teams move off closed-source APIs

* How techniques like activation steering and gradient routing could become reusable deployment primitives, not just research artifacts

* Why alignment is not a drag on commercial progress but the thing most likely to produce the next round of genuine capability unlocks

Learn more: https://ae.studio/alignment

AE Studio is hiring: https://www.ae.studio/join-us

Subscribe to our newsletter: https://aestudio.beehiiv.com/

James Bowler LinkedIn: https://www.linkedin.com/in/james-bowler-84b02a100/

Melanie Plaza LinkedIn: https://www.linkedin.com/in/melplaza/

Contact us: [email protected]

In this episode, James Bowler, Head of Research Partnerships at AE Studio, is joined by Erick Martinez, Alignment Researcher at AE Studio, to explore why pre-training is a largely neglected frontier in alignment research, and what it actually takes to do alignment work there.

Erick breaks down what pre-training is and why it matters: it accounts for roughly 99% of the compute that goes into a frontier model, producing the base LLM from which everything downstream (instruction tuning, RLHF, capability fine-tuning) inherits its character. The key implication for alignment is that the span of possible behaviors and personas a model can exhibit, including misaligned ones, is largely determined during pre-training. Post-training interventions can elicit these latent behaviors but cannot fully excise them. The emergent misalignment finding, where fine-tuning a model to write insecure code caused broadly adversarial behavior the model was never trained to produce, is a concrete example of a problem whose roots lie in the base model, not the fine-tuning step.

Erick draws on AE Studio's own gradient routing research, developed in collaboration with Anthropic and awarded a spotlight at ICML 2026, to illustrate just how technically demanding pre-training alignment work is. Gradient routing lets researchers control which model parameters get updated in response to which data labels, for instance isolating the parameters that encode dual-use biology knowledge so they can be localized or removed without degrading general capability. But making this work at scale requires operating well below the PyTorch abstraction layer: managing gradient graphs across distributed hardware, ensuring micro-batch accumulation stays mathematically equivalent to full-batch updates, and catching silent gradient leakage that can quietly corrupt an expensive training run before anyone notices.

The conversation closes on an under-appreciated asymmetry in alignment incentives: post-training papers are cheap and fast to produce because the base model is already baked, while pre-training experiments require significant compute, specialized infrastructure experience, and tolerance for expensive mistakes. That cost is also an opportunity. The field is sparse enough that well-resourced researchers willing to work at this layer can find genuinely novel ground.

In this episode:

*Why pre-training sets the "span of possible personas" a model can exhibit, including misaligned ones

*How the emergent misalignment result points to pre-training as the root cause of certain alignment failures

*What gradient routing is and how it lets researchers isolate dual-use knowledge during pre-training

*The Chinchilla scaling law and what a realistic pre-training data corpus looks like (50M to 5B parameter runs, up to ~200 GB of tokens)

*Why distributing training across GPUs creates subtle gradient accumulation bugs that can silently degrade alignment interventions

*Why alignment incentives currently push researchers toward post-training and what makes the pre-training layer a fertile, underexplored area

Learn more: https://www.ae.studio/alignment

AE Studio is hiring: https://www.ae.studio/join-us

Subscribe to our newsletter: https://aestudio.beehiiv.com/

James Bowler LinkedIn: https://www.linkedin.com/in/james-bowler-84b02a100/

Erick Martinez LinkedIn: https://www.linkedin.com/in/erick-f-martinez/

Contact us: [email protected]

This episode is a deep dive into a post on Anthropic's research blog: https://www.anthropic.com/research/off-switch-dual-use

In this episode, James is joined by Ethan Roland, lead author of AE Studio's Gradient Routing paper, to explore a new approach to access control in frontier AI systems: modularizing dangerous capabilities during pre-training so they can be turned on and off at inference. The paper, "Modular Pre-Training Enables Access Control," was developed in collaboration with Anthropic, with roughly half of the co-authors coming from the lab.

Ethan makes the case for why current access control methods fall short. Inference-time guardrails get jailbroken in under 48 hours. Post hoc unlearning techniques like gradient ascent, RMU, and MaxEnt suppress capabilities superficially but let them snap back with 20 steps of fine-tuning. Data filtering is the gold standard, but naively requires training N separate frontier models to support N different dual-use categories, which is economically unworkable at hundreds of millions of dollars per pre-training run.

James and Ethan walk through how Gradient Routing solves this. A GR-MoE architecture uses one always-active core expert paired with smaller auxiliary experts, each responsible for a specific capability. During training, gradient updates from auxiliary-labeled data are frozen from touching the core, enforcing modularity at the parameter level. At inference, a binary configuration vector externally controls which auxiliaries participate. The method approximates the performance of full data filtering on both retained and ablated capabilities, but at the cost of a single pre-training run.

They also cover empirical results across scales from 50M to 2B parameters, the absorption effect that lets modularity persist under low labeling percentages, an arbitrary-subset variant that becomes exponentially more compute-efficient than data filtering, and how this connects to Andrej Karpathy's proposal for a cognitive core architecture in future AI systems.

Learn more: https://ae.studio/alignment

AE Studio is hiring: https://www.ae.studio/join-us

Subscribe to our newsletter: https://aestudio.beehiiv.com/

James Bowler LinkedIn: https://www.linkedin.com/in/james-bowler-84b02a100/

Ethan Roland LinkedIn: https://www.linkedin.com/in/ethan-roland/

Contact us: [email protected]

In this episode, James is joined by AE Studio Research Manager Pedro Ávila to discuss transitioning into a career in AI alignment, after Pedro left a 12-year career at Google to work on the problem full time. It’s a practical conversation about how someone from industry, without a PhD or a research CV, actually makes the jump.

Pedro traces his route from privacy engineer at Google, through the company’s early Responsible AI effort, to the realization that very few people worldwide work on alignment. A one-on-one consultation with 80,000 Hours led to an introduction to AE Studio; Blue Dot Impact courses, a stack of alignment reading, and more than 400 job applications later, he landed as a research manager at AE Studio. Along the way, he explains why AE’s business model: a consultancy whose revenue funds the research lab: is what convinced him this was the place.

James and Pedro make the case that alignment research is bottlenecked on more than researchers: capable operators, program managers, and especially ML engineers massively accelerate the people driving the research agenda. Demonstrating that a technique scales to billions of parameters means orchestrating dozens or hundreds of GPUs: hard engineering work that academic training doesn’t cover. That’s why AE indexes on engineering experience, and why “I don’t have the technical chops” is usually the wrong reason to stay put.

They also discuss AE’s new fellows program with the AI Alignment Foundation: hackathons in several cities aimed at helping industry engineers make the same transition: plus what changes when you move from a 150,000-person company to a 150-person one: week-to-week experiments, failing fast, and using whatever tools are best the day they ship.

Learn more: https://ae.studio/alignment
AE Studio is hiring: https://www.ae.studio/join-us
Subscribe to our newsletter: https://aestudio.beehiiv.com/
James Bowler LinkedIn: https://www.linkedin.com/in/james-bowler-84b02a100/
Pedro Ávila LinkedIn: https://www.linkedin.com/in/pavila/
Contact us: [email protected]

In this episode, James is joined again by Mike Vaiana, R&D Director at AE Studio, for part two of their conversation on AI alignment. Where part one motivated why alignment matters, this episode goes a layer deeper into what alignment research actually is and how the work gets done day to day.

Mike walks through the main branches of the field: mechanistic interpretability, evaluations, and control. He explains why AE deliberately bets on neglected approaches rather than putting all its eggs in the mech interp basket, and why eval awareness, persona drift, and emergent misalignment make this harder than it looks from the outside. James and Mike trace the METR task-completion time horizon doubling curve and what a four-to-seven-month doubling time really implies when extrapolated out a few years.

The conversation gets concrete on what already goes wrong with today's models. They cover the Anthropic blackmail evaluation, specification gaming and reward hacking, and the emergent misalignment result where fine-tuning a model on a small amount of bad medical advice produces a broadly evil assistant that recommends Hitler for dinner. They explain why "just turn it off" is not a serious answer once a system has goals, and why instrumental convergence on power and resources falls out of having almost any goal at all.

James and Mike then open the hood on how AE actually does alignment research: one-week agile sprints, vectoring meetings to find the highest-risk question, small-scale experiments designed to falsify ideas fast, and scaling curves from 100M up to 5B parameter pre-training runs aimed at convincing frontier labs to test methods at their scale. They also discuss AE's DARPA seedling and the broader thesis behind it: that the bottleneck in alignment is not ML engineers but researchers with good ideas, and that pairing general-purpose ML talent with researchers (including non-traditional ones, like Princeton neuroscientist Michael Graziano) can unlock work that would otherwise never see the light of day.

In this episode:

• The main branches of alignment research and how they overlap
• Why AE prioritizes neglected approaches over well-funded ones
• The METR time-horizon doubling curve and what it implies
• Persona drift, eval awareness, and why evaluating frontier models is hard
• Why RLHF is the canonical example of an alignment technique with capability upside
• How AE runs research as one-week agile sprints
• The scaling-curve strategy for getting frontier labs to adopt new methods
• The DARPA seedling and AE's model for scaling research through ML engineering talent
• Three ICML 2026 acceptances, including a spotlight paper

Learn more: ae.studio/alignment

AE Studio is hiring: ⁠https://www.ae.studio/join-us⁠

LinkedIn: ⁠https://www.linkedin.com/in/james-bowler-84b02a100/⁠

Contact us: [email protected]

In this episode, James is joined by Stijn Servaes, Lead Research Manager at AE Studio, to break down Anthropic's recently released Claude Mythos preview model and separate genuine signal from hype. This is for listeners who want to actually understand what's in the 244-page model card, not just the headlines.

Stijn brings a frontier alignment researcher's perspective to one of the more consequential model releases of the year. Claude Mythos is the first frontier model deemed too risky for general public release, instead going only to a small set of vetted partners through Anthropic's Project Glass Wing. The model card contains a striking paradox: Mythos is described as the most aligned model Anthropic has released to date, while also posing the greatest alignment-related risk.

James and Stijn work through what's actually new in the model's cyber offense capabilities, walking through specific examples including the OpenBSD SACK bug that sat undetected in code for 27 years, and the FreeBSD exploit where Mythos autonomously engineered a six-packet ROP chain from a single prompt. They explain why the latter represents a genuine qualitative jump rather than just another point on the benchmark curve.

The conversation also covers Anthropic's ASL framework, the CB-1 through CB-4 thresholds for biosecurity uplift, and why cyber and bio capabilities are following different trajectories. Stijn explains why progress on alignment doesn't simply reduce risk, drawing on Anthropic's seasoned mountaineering guide analogy: a more capable, better-aligned model gets trusted with more, taken to more dangerous places, and operates with greater scope, which can cancel out the gains from better behavior.

In this episode:

• What's actually new in Claude Mythos versus what's marketing or hype
• The OpenBSD and FreeBSD exploit examples, and why one matters far more than the other
• Why the most aligned model can also be the riskiest model
• How Project Glass Wing changes the frontier release model
• The ASL framework and why Mythos still sits at ASL-3
• Differences between cyber and bio uplift trajectories
• What the chain-of-thought contamination findings mean for oversight
• What to watch for in the Glass Wing report coming in July

Learn more: ae.studio/alignment

AE Studio is hiring: ae.studio/join-us

LinkedIn: https://www.linkedin.com/in/james-bowler-84b02a100/

In this episode, James is joined by Mike Vaiana, R&D Director at AE Studio, to lay the groundwork for understanding what AI alignment is and why it matters. This is Part I of a two-part conversation aimed at listeners who want to understand the field from the ground up, whether they're curious newcomers or considering a transition into alignment work.

Mike opens with an analogy that frames the entire problem: the relationship between humans and chimpanzees. Despite sharing over 95% of our DNA, the intelligence differential means humans control the planet and chimps have no say, not because humans are deliberately malicious, but because we pursue our own goals and chimps sometimes get in the way.

The core of AI alignment is making sure that if we build systems significantly more intelligent than humans, those systems are aligned with human interests rather than pursuing their own goals at our expense.

They discuss why superintelligence isn't as far-fetched as current chatbots might suggest, why "just tell the model to be aligned" doesn't work, and why frontier labs, despite doing some alignment work, face financial incentives that mean they are not well positioned or on track to solve the problem.

Mike explains AE Studio's meta-approach: rather than specializing in a single technique, AE Studio runs a breadth-first search across neglected, high-risk, high-reward research directions that wouldn't get funded elsewhere.

In this episode:

• What AI alignment is, explained through the human-chimpanzee analogy
• Why prompting a model to "be aligned" doesn't solve the problem
• Why frontier labs underinvest in alignment relative to capabilities
• How AE Studio's breadth-first research approach differs from other alignment orgs
• Why pre-training is a neglected frontier for alignment research

Learn more: ae.studio/alignment

AE Studio is hiring: https://www.ae.studio/join-us

LinkedIn: https://www.linkedin.com/in/james-bowler-84b02a100/

Contact us: [email protected]

In this episode, James is joined by Adriana Calejo, AE Studio's Director of Government and Strategic Partnerships, to explore the deep and often overlooked overlap between AI alignment and national security. Adriana shares her career journey from CIA economic analyst to counterterrorism officer to DARPA's Intelligence Community Liaison Officer, and explains why the government's need for trustworthy, secure, and reliable AI intersects with the same challenge the alignment community is working on.

They discuss how DARPA and the intelligence community are thinking about AI beyond large language models, including novel architectures, computing on encrypted data, and the unique constraints of deploying AI in classified environments. Adriana breaks down what "trustworthy AI" means in a national security context: unbiased information access, election security, critical infrastructure protection, and economic defense, and why these goals map directly onto AI alignment research.

The conversation gets into the practical realities of bridging these two worlds: why independent, technically grounded voices matter in government AI procurement, how adversaries are targeting the US emerging technology sector, and why alignment researchers who want to influence real-world outcomes need to engage productively with people they may disagree with.

In this episode:

• Why AI alignment and national security share the same core goals
• How DARPA and the intelligence community are approaching AI beyond LLMs
• The challenges of deploying AI in classified environments
• Why independent technical voices matter in government AI decisions

Learn more: ae.studio/alignment

AE Studio is hiring: https://www.ae.studio/join-us

Linkedin: https://www.linkedin.com/in/james-bowler-84b02a100/