Inference by Turing Post

Regarding the question of what matters most – the model or the harness – Michael Bolin is somewhere in the middle.
Stronger models clearly pushed Codex to new heights. But without the right harness around them, those models would not be able to operate reliably, and – most importantly – safely on a real developer’s machine. At least, not yet.
In this episode of Inference, I talk with Michael Bolin – lead for open source Codex at OpenAI – about the engineering layer that makes coding agents actually function: the agent loop, sandboxing, tool orchestration, and the design decisions that determine how much freedom an agent should have.
In this conversation, we get into:
What a harness actually is and why every coding agent needs one
Can a model be enough for a reliable coding workflow
Why do they build harness as small and tight as possible
How Codex handles sandboxing and security across OS
Why safety and security are not the same thing in agentic systems
How coding agents are changing the daily workflow of developers
Why documentation, tests, repo structure, and agents.md suddenly matter more
Whether too much context can make an agent worse
Why Michael believes the future may involve fewer tools, but more powerful ones
If you’re trying to understand where coding agents are actually going, this episode is for you.
Subscribe to the channel to be notified about Part 2, where we discuss what becomes of the software engineer in the age of agents.
Chapters:
0:00 The New Inner Loop of AI Coding Agents
0:17 Introduction: Michael Bolin and Open Source Codex
1:17 What the “Harness” Is in AI Coding Agents
2:13 Security and Sandboxing for AI Agents
4:33 Codex Launch and Rapid Growth
5:25 The Codex App: A New Interface for Developers
6:36 How Coding Agents Change Developer Workflows
10:04 Writing Codebases and Documentation for AI Agents
12:44 Context Engineering and Prompting for Codex
16:02 Model vs Harness: What Really Matters for Agents
19:23 Multi-Agent Systems, Tools, and the Future of AI Development
*Follow on*: https://www.turingpost.com/
*Did you like the episode? You know the drill:*
📌 Subscribe here and here (https://www.turingpost.com/subscribe) for more conversations with the builders shaping real-world AI.
💬 Leave a comment
👍 Like it
🫶 Thank you for watching and sharing!
*Guest:*
 Michael Bolin, tech lead on Codex, OpenAI
https://www.linkedin.com/in/michael-bolin-7632712/
https://x.com/bolinfest
https://github.com/openai/codex
📰 Transcript: https://www.turingpost.com/bolin1
*Turing Post* – AI stories from labs the Valley doesn't cover.
https://x.com/TheTuringPost
https://www.linkedin.com/in/ksenia-se
Tags: #AI #OpenAI #Codex #CodingAgents #DeveloperTools #AgenticAI #SoftwareEngineering #HarnessEngineering #Harness

In this episode, Ioannis Antonoglou, co-founder and CTO @ReflectionAI (ex-DeepMind, AlphaGo/AlphaZero/MuZero) explains what they are building: a frontier open-weight “general agent model” trained end-to-end with pretraining plus reinforcement learning.
And I’ll be honest: I left this conversation more skeptical than I expected. They raised $2 billion last year. But where the results?
Reflection’s thesis is huge – build the missing Western open base model, then use RL to push it to the frontier. The problem is that this is also the slowest path in the game. “All hands on deck building the model” means no clear wedge product yet, few concrete proof points, and a lot of execution risk while closed labs keep shipping.
Am I missing something? Watch the video and leave your opinion in the comments
Chapters:
0:00 Building AGI and the Mission Behind Reflection
0:25 From AlphaGo to Today: How AI Progress Really Happens
2:11 Breakthroughs vs. Engineering: What Still Matters Most
3:10 Defining AGI and Why It May Not Need Huge Breakthroughs
3:41 Why Reflection Shifted from Coding Agents to Frontier Models
5:15 The New Focus: Open Frontier Models and General Agents
6:33 Bottlenecks in Building Frontier AI: Team, Compute, and Scale
7:48 AI Tools, Internal Workflows, and Model-First Strategy
8:24 Can Open Models Catch Closed Labs?
10:34 Reinforcement Learning, Research Priorities, and Advice for Young Builders
14:01 Joining DeepMind, Open Science, and the Book That Shaped Him
*Follow on*: https://www.turingpost.com/
*Did you like the episode? You know the drill:*
📌 Subscribe for more conversations with the builders shaping real-world AI.
💬 Leave a comment if this resonated.
👍 Like it if you liked it.
🫶 Thank you for watching and sharing!
*Guest:*
Ioannis Antonoglou, Co-Founder, President & CTO at Reflection AI https://x.com/real_ioannis https://www.linkedin.com/in/ioannis-alexandros-antonoglou-45393253 https://reflection.ai/
📰 Transcript: https://www.turingpost.com/nathan
*Turing Post* – AI stories from labs the Valley doesn't cover.
https://x.com/TheTuringPost
https://www.linkedin.com/in/ksenia-se
Tags: #reflectionai #opensource #deepmind #ai #openclaw #aisafety

Ioannis Antonoglou helped build AlphaGo, AlphaZero, and MuZero at DeepMind. Now he’s CTO and co-founder of Reflection AI, betting that frontier models should be open weights, not a black box behind an API.
In Part 1, we talk about openness as an actual strategy: why open models can move faster, why “sovereignty” matters for enterprises and governments, and why safety might improve when the ecosystem can stress-test the system instead of guessing.
We also get into the uncomfortable part: capable open agents can misbehave in public, fast (OpenClaw is the recent reminder). Is that a reason to close everything up, or a reason to make the risks visible and fixable?
Topics covered:
 – Why a former DeepMind builder chose open weights
 – Open models as a commercial engine (and what investors bought)
 – Openness, safety, and “more eyes on the system”
 – Concentration of AI power in closed labs
 – Who open frontier models are really for (research, enterprises, governments)
Subscribe for Part 2: how Reflection plans to compete with closed labs and what they’re building under the hood.
Chapters:
0:00 — “No One Was Sharing This Information”
0:16 — From DeepMind to Reflection AI
0:52 — Why Move from Closed Labs to Open Weights?
2:20 — Pitching Open Models Before the DeepSeek Moment
3:31 — What Changed in the Past Year
4:43 — Why Openness Accelerates Scientific Progress
6:06 — Open Source vs Safety: The Open Claw Case
7:19 — The Real Concern: Concentration of AI Power
8:23 — The Open Source Paradox
9:11 — The Value Proposition of an Open Frontier Model
*Follow on*: https://www.turingpost.com/
*Did you like the episode? You know the drill:*
📌 Subscribe for more conversations with the builders shaping real-world AI.
💬 Leave a comment if this resonated.
👍 Like it if you liked it.
🫶 Thank you for watching and sharing!
*Guest:*
Ioannis Antonoglou, Co-Founder, President & CTO at Reflection AI https://x.com/real_ioannis https://www.linkedin.com/in/ioannis-alexandros-antonoglou-45393253/ https://reflection.ai/
📰 Transcript: https://www.turingpost.com/antonoglou_part1
*Turing Post* – AI stories from labs the Valley doesn't cover.
https://x.com/TheTuringPost
https://www.linkedin.com/in/ksenia-se
Tags: #reflectionai #opensource #deepmind #ai #openclaw #aisafety

Open models are often discussed as if they’re competing head-to-head with frontier systems. Are they catching up? Falling behind? Are they “good enough” yet?
Nathan Lambert doesn’t believe open models will ever catch up with closed ones, and he explains clearly why. But he also argues that this is the wrong framing.
Nathan is a research scientist at the Allen Institute for AI, the author of the RLHF Book, and the writer behind the Interconnects newsletter. He’s also one of the clearest voices on what open models are for, and just as importantly, what they are not.
We talk about how academic AI research lost influence as training scaled up, why open models became the main place where experimentation still happens, and why that role matters even when open models trail frontier systems. We also discuss why China’s open model ecosystem developed so differently from the US one, and what that tells us about incentives, talent, and access to resources.
From there, the conversation moves into the mechanics: post-training and reinforcement learning complexity, data availability, coding agents, hybrid architectures, and the very practical reasons most people continue to rely on closed models, even when they support openness in principle.
This is a conversation about how AI research actually moves, where open models fit into that picture, and what it means to build systems when the frontier is expensive, fast-moving, and increasingly product-driven.

This conversation offers a realistic look at where the open ecosystem stands today. Watch it!
*Follow on*: https://www.turingpost.com/
*Did you like the episode? You know the drill:*
📌 Subscribe for more conversations with the builders shaping real-world AI.
💬 Leave a comment if this resonated.
👍 Like it if you liked it.
🫶 Thank you for watching and sharing!
*Guest:*
Nathan Lambert, Research Scientist at Allen Institute for AI (AI2)
https://x.com/natolambert
https://www.linkedin.com/in/natolambert/
https://www.interconnects.ai/ (his newsletter on open models + RL + everything important in AI)
https://rlhfbook.com/ - The RLHF Book
https://allenai.org/
*Links:*
State of AI in 2026 (Lex Fridman interview): https://www.youtube.com/watch?v=EV7WhVT270Q&t=10206s
NVIDIA’s path to open models https://www.youtube.com/watch?v=Y3Vb6ecvfpU
OLMo models: https://allenai.org/olmo
NVIDIA Nemotron: https://developer.nvidia.com/nemotron
SpaceX + xAI partnership: https://www.spacex.com/updates#xai-joins-spacex
Season of the Witch (book): https://www.simonandschuster.com/books/Season-of-the-Witch/David-Talbot/9781439108246
📰 Transcript: https://www.turingpost.com/nathanlambert
*Turing Post* – AI stories from labs the Valley doesn't cover.
https://x.com/TheTuringPost
https://www.linkedin.com/in/ksenia-se

First Western interview with a senior MiniMax researcher. Olive Song explains how they actually build models that work.
When MiniMax's RL training wouldn't converge, they debugged layer by layer until they found it: fp32 precision in the LM head. When their models learned to "hack" during training, exploiting loopholes to maximize rewards, they had to rethink alignment from scratch. When benchmarks said their models were good but production said otherwise, they discovered the problem: environment adaptation.
Olive talks about working at a pace where new models drop at midnight and you test them at midnight. How they use an internal AI agent to read every new paper published overnight. Why they sit with developers during experiments to catch dangerous behaviors in real-time. What "ICU in the morning, KTV at night" means when results swing wildly. How problem-solving becomes discovery when you're debugging behaviors no one has seen before.
This is how Chinese labs are moving fast: first-principles thinking, engineering discipline, and willingness to work whenever the model in experimentation requires you to.
We spoke on Sunday at 9 pm Beijing time. Olive was still waiting for results from new model experiments, so my first question was obvious: does everyone at the company work like this?
*Follow on*: https://www.turingpost.com/
*Did you like the episode? You know the drill:*
📌 Subscribe for more conversations with the builders shaping real-world AI.
💬 Leave a comment if this resonated.
👍 Like it if you liked it.
🫶 Thank you for watching and sharing!
*Guest:* Olive Song, Senior Researcher at MiniMax
MiniMax: https://www.minimaxi.com/
Models: https://huggingface.co/MiniMaxAI
*Links:*
vLLM: https://github.com/vllm-project/vllm
SGLang: https://github.com/sgl-project/sglang
📰 Transcript: https://www.turingpost.com/olive
Chapters:
0:00 – Reinforcement Learning and Unexpected Model Behaviors
3:08 – Roleplay, Alignment, and “AI with Everyone”
4:02 – How AI Changes Daily Life and Productivity
4:59 – Inside Miniax: How Researchers and Engineers Work Together
5:32 – Human Alignment and Safety in Open Models
6:16 – Why Engineering Details Matter More Than Algorithms
8:17 – Open Weights: Benefits, Risks, and Responsibility
10:57 – Specialization vs General AI Models
12:07 – Agentic AI and Long-Horizon Tasks
29:50 – AGI, Creativity, and the Future of AI
*Turing Post* – AI stories from labs the Valley doesn't cover.
https://x.com/TheTuringPost
https://www.linkedin.com/in/ksenia-se
#MiniMax #ReinforcementLearning #AIResearch #OpenWeights #ChineseAI #OpensourceAI