Conversation With Baseten's Head of Training

Charles O'Neill, Baseten's co-Head of Training and a Parsed co-founder, explains how he entered AI research, why Parsed bet on specialized open-weight models, and how post-training and inference change who can control a useful model.

He expects closed frontier labs to retain a scale advantage. His narrower claim is that a smaller model trained on one company's task, data, and feedback can beat a general system on that task while running faster and cheaper. He also accepts short-term misuse and job-displacement risks from wider access.

• Guest: Charles O'Neill
• Interviewer: Madison
• Publisher: Baseten
• Published: July 14, 2026
• Source: Original video

Parsed began in London near Oxford. After four days in San Francisco, Charlie thought the team had learned more than it had during the previous four or five months. Customers, researchers, events, and chance meetings gave the founders more information about where the field was moving. The eight- or nine-person team relocated within weeks and lived together.

That timing and headcount come from his recollection. Research on technical clusters supports local knowledge spillovers, but also finds higher labor and housing costs in dense hubs (NBER). It cannot show that the move caused Parsed's acceleration.

Baseten acquired Parsed in December 2025. Its announcement says the deal joined Parsed's training systems with Baseten's inference stack.

Charlie had planned to begin a PhD at Oxford. He left because language models had increased what a small team could attempt. A narrow academic contribution looked less attractive than taking a larger risk on a problem with direct impact.

He wanted the company to keep a research habit: choose problems on a longer horizon. Startups can let immediate revenue decide every move. Parsed expected basic post-training services to commoditize and chose harder, less-verifiable work in medicine, insurance, and legal systems.

Later company material matches that positioning. It does not prove that research training produces better startup decisions or that competitors ignored those domains.

Charlie recommends learning from a working system downward. Run an open reinforcement-learning framework, observe a learning curve, change a parameter, then study the code and mathematics exposed by the experiment. fast.ai teaches a similar project-first sequence.

This does not remove the need for knowledge. A researcher still needs enough context to judge which problems matter and which ideas can work. Charlie calls that a value function over research directions.

His example starts with specialized continual learning. Compress a growing KV cache, train with compaction in the loop, let a model invoke compaction itself, then investigate whether frequently reused state can move into adapters or model weights. Baseten's STILL work reports bounded KV-cache compression. Iterative compaction and consolidation into weights remain proposed steps.

Charlie thinks unusual experience can change how a researcher sees a problem. His background was philosophy and literature. It made him comfortable treating language models as messy behavioral systems instead of deterministic logic circuits.

Reading the early GPT work pulled him toward mathematics and computer science. GPT-2's improvement over the first GPT convinced him that more data and model scale could keep producing gains. The published GPT-2 report and later scaling-law work support that technical history. They do not establish the broader claim that a distinctive background causes better research.

His advice is to find the experience or skill that changes how you work, then use it during actual research instead of treating paper-reading as the whole job.

Max Kirkby introduced Charlie to Mudith Jayasekara at an Oxford pub. Charlie says the three decided to build together after one short meeting. Their first attempt was a narrow hospital AI pitch designed for investors. They abandoned it hours before the meeting because it did not match their interests or strengths.

They pitched the team and a research direction instead. Charlie is unsure whether the meeting was with Lightspeed. Max's later account names LocalGlobe as the seed lead.

Parsed then committed to post-training open-weight models for specific jobs. Task specialization can work: the LoRA Land study found large in-domain gains from small tuned models. Other work finds that strong in-domain judges can still lose on generalization and fairness. A narrow win is not broad model superiority.

“Big Token” is Charlie's label for a future in which one or two labs control useful AI. He does not advise teams to reject those labs on day one. Start with the strongest hosted model, use prompting and scaffolding to prove the task, then measure the ceiling.

Move toward an open-weight model when proprietary data, feedback, cost, latency, or control justifies the work. Owning weights lets a team choose training methods and deployment infrastructure. Closed providers also offer fine-tuning, so the difference is degree of control rather than prompting versus training.

Open-weight is the more precise term for many models in this discussion. The Open Source Initiative requires broader freedom and access than downloadable weights alone.

Charlie also warns that a product built entirely around a frontier API can lose its differentiation when the provider adds the same workflow. The response is not a wrapper with more prompts. It is a learning loop built from the customer's own task and feedback.

Pre-training teaches next-token prediction over a large corpus. Post-training then shapes behavior through continued training, supervised examples, preference data, reinforcement learning, distillation, and other methods. Charlie uses “fine-tuning,” “post-training,” and “specialization” loosely, though they are not strict synonyms.

Reinforcement learning works best when a task has a checkable result. Code can run unit tests. A formal proof can compile. The model samples an attempt and receives feedback on the outcome. That signal is closer to the model's own behavior than supervised imitation, but it is sparse and creates a credit-assignment problem.

InstructGPT combined supervised demonstrations with reinforcement learning from preferences. DeepSeek-R1 showed that large-scale RL could elicit reasoning, while R1-Zero's readability and language-mixing failures required cold-start data and multi-stage training. RL alone did not produce the final system.

Specialization can improve a bounded task and reduce serving cost. It also adds evaluation, training, quantization, kernel, and hosting work. Charlie therefore returns to the strongest hosted model first: understand the task, build an evaluation, then work down the model ladder.

Charlie expects closed labs to keep an intermittent base-model advantage because they get more capital, compute, and early access to new hardware. He thinks open-weight models can remain economically important without leading every benchmark. A company can tune one to its own distribution while a closed general model serves many unrelated users.

The lag is small but hard to reduce to one number. Epoch AI estimated a four-month average gap in May 2026 and warned that benchmark coverage may understate it. Other methods produce different results.

Compute is only one constraint. Training needs model size, compute, and data in balance. Proprietary task data and real-user feedback may become more valuable as public text becomes less useful at the margin. Forecasts of data exhaustion remain conditional on synthetic data, multimodal sources, transfer, and efficiency gains.

Kimi K2 shows why parameter counts need care. It has one trillion total parameters but activates 32 billion per token. Total parameters across dense and mixture-of-experts models do not directly measure capability or runtime cost.

Charlie expects some skilled cognitive work to disappear. He rejects the assumption that demand stays fixed and every automated task becomes a permanent job loss. Earlier technologies also created new products and occupations.

Inference engineering is his example. Mixture-of-experts routing, KV-cache management, batching, networking, and model-specific kernels create difficult systems work. That does not prove net job growth.

The ILO's 2025 task-level study found that one in four workers had some generative-AI exposure and judged transformation more likely than full redundancy. Its 2026 review also warns about entry-level opportunities, inequality, autonomy, and job quality. U.S. projections show the split directly: software-developer employment rises while computer-programmer employment falls.

Start by serving a model through vLLM, TensorRT-LLM, or SGLang. Measure the workload before studying every low-level component. If prefill dominates, inspect context handling and KV-cache movement. If decode or throughput dominates, examine batching, draft models, kernels, memory, and networking. The major engines expose high-level APIs while retaining controls for these layers.

Charlie says Parsed chose Baseten as its sole inference provider before the acquisition. Joining combined Parsed's training work with serving infrastructure and removed its weak position as an intermediary between customers and GPU providers. Baseten's announcement supports that broad account; comparative performance and research-autonomy claims remain his.

The final exchange turns to control and safety. Charlie accepts that downloadable models can increase short-term misuse, including cyber misuse. He fears concentrated control more and argues that wider access also lets defenders adapt. The International AI Safety Report 2026 identifies both sides: open weights aid research and customization, but safeguards are easier to remove and a release cannot be recalled.

Compute still limits access. Charlie distinguishes expensive training from low-volume serving, but overstates how little hardware the largest models need. Moonshot's Kimi K2 deployment guide requires at least 16 H200 or H20 GPUs for its mainstream FP8 setup, not one eight-GPU node. Precision, context length, traffic, and hardware determine whether a model fits and serves well.