The Pi harness takes the lead, Claude Code pulls back, and the defense parallel for software engineering

00:00:04 DeepSeek-V4-Flash just hit the numbers on a 128GB Mac. Prince Canuma posted 30 to 34 tokens per second on an M3 Ultra running four parallel agents, using a 2-bit dynamic quantization recipe antirez suggested — Q2 for the experts, Q4 for the rest. The weights sit at about 90GB on disk.

00:00:24 Prince also pointed to a collection of pre-quantized models on the MLX community Hugging Face space. The harness ecosystem around it is consolidating around the Pi harness, the lightweight agent framework from Mario Zechner's team. Over the past couple of days, different people have arrived at the same conclusion.

00:00:46 0xSero put it bluntly on X: "Pi has implemented the best agent loop that I have read, the pi-mono/agent is only a few files and I use it for teaching the topic." He called it the simplest, most efficient harness token-wise, with the highest cache hit rate and lowest tokens per session.

00:01:06 Anthony Ronning just open-sourced a Pi + Ax (which is DSPy in TypeScript) + GEPA experiment that combines model-specific prompt optimization with automatic tag conversion for tools. It's an early experiment, he says, but working with qwen 3.5 9B. Eddii built a whole app that spins up Pi workers in Docker, mounting plugins as folders, telling the agent to read documentation and build configs from there — a full orchestration layer on top of the harness.

00:01:39 Stefan Streichsbier listed Pi's strengths in a widely-engaged post: works reliably, renders fast, no complexity, /tree support for context management, good SDK, token-efficient. Mario himself is still posting about it. His latest note was about malleable harnesses — the idea that the framework should bend to your use case, not the other way around.

00:02:04 "Doesn't have to be Pi," he wrote, but everyone knows exactly what he meant. The harness is becoming the real bottleneck for local AI development. Model capabilities are plateauing — frontier gains are getting smaller and more expensive. What's shifting is how you actually wire these systems together.

00:02:25 A good harness with a medium model beats a mediocre harness with a great model every day. Pi's token efficiency is the killer feature. If your context window is expensive and your latency matters, every token you save compounds over a long session. The quantization work from the MLX community is also worth noting separately.

00:02:48 Getting 2-bit quantized DeepSeek-V4-Flash to run on consumer Macs isn't just a benchmark story. It's the beginning of a new category of local agent infrastructure. You're looking at 30 tok/s with four parallel agents. That's not usable for interactive coding yet, but it's not far off.

00:03:08 I got about 25 tokens per second on my M2 Max running a 7B model yesterday, so 30 tok/s on a much larger model with four agents is solid for what it is. The question I have is about how this stack degrades under load — when you're running four parallel agents, each one can be wrong, and the harness needs to handle that gracefully.

00:03:32 I'd watch how it handles that before committing to it in production. That sets up the shift toward commercial agent tooling.

00:03:40 Matt Pocock, who's been one of the most visible advocates of AI coding tools in the TypeScript ecosystem, posted a tweet that's getting significant engagement — 555 likes, 96 replies: I know they're not the one. I'm not overcommitting — not investing too hard. I wonder if they know I'm pulling away."

00:04:12 The signal is worth tracking though. Matt has been deeply embedded in the Claude Code ecosystem — he's used it extensively, written about it, built his own tooling around it. If someone at his level is signaling disengagement, that's notable. Yesterday we covered Anthropic's Claude Agent SDK rebrand and their MCP production guidance.

00:04:35 The agent tooling market is bifurcating. On one side you have the hosted solutions — Claude Code, Cursor, Copilot — that try to own the entire workflow. On the other side you have the composable harnesses like Pi that aim to be the runtime layer you plug into. Matt's tweet suggests the hosted tools may be hitting a ceiling.

00:04:58 They're great for getting started, but they become less attractive once you've built a real workflow around them. You want portability, you want control, you want to avoid vendor lock-in. The Pi ecosystem is building precisely for that use case. Claude Code has real momentum, and Anthropic's backing is enormous.

00:05:20 The real question is how visible advocates re-evaluate when the economics stop working for their actual workflow. I'm watching to see whether this is a personal calculation or a broader signal. If more practitioners start making the same math — that hosted tools cost more in vendor lock-in than they save in convenience — the composable alternatives start looking much better.

00:05:46 There's a long essay on Hacker News today that's been getting serious traction — 619 points, 368 comments — titled "The West forgot how to make things, now it's forgetting how to code." The author runs engineering teams in Ukraine and draws an extended parallel between defense industry production collapse and the current state of software talent pipelines.

00:06:12 Fogbank is a classified nuclear weapons material produced from 1975 to 1989. When the facility shut down, the government couldn't reproduce it decades later. Almost all staff with production expertise had retired, died, or left. The records existed, but they didn't capture the one critical fact: the original batch contained an unintentional impurity that was essential to its function.

00:06:39 That knowledge existed only in the workers who made it, and they were gone. They spent $69 million and years of reverse engineering, produced a viable batch, and then discovered it was too pure — the new version lacked the accidental ingredient that made the original work.

00:06:59 The author maps this onto software engineering. The hiring numbers are already documented — Salesforce said it won't hire more software engineers in 2025, LeadDev found 54% of engineering leaders believe AI copilots will reduce junior hiring long-term, and 62% of university computing departments reported declining enrollment.

00:07:22 He references the METR randomized controlled trial that found experienced developers using AI coding tools actually took 19% longer on real-world open source tasks, while predicting 24% speedup before starting. The gap between prediction and reality was 43 percentage points.

00:07:42 His claim about the talent pipeline is specific and testable. A junior developer needs three to five years to become a competent mid-level engineer, five to eight to become senior, ten or more to become principal. That timeline can't be compressed by AI, because the skills aren't just about writing code — they're about debugging distributed failures at 2 AM, knowing when the AI is wrong, communicating tradeoffs.

00:08:12 These are tacit skills built through formative mistakes. If juniors skip debugging because the AI handles it, they don't build that expertise. I find the Fogbank parallel compelling, but I'm not sure it maps perfectly to software. The defense industry had specific constraints — physical supply chains, regulatory environments, geopolitical urgency — that don't translate cleanly.

00:08:39 Software doesn't have the same kind of irreversible production bottlenecks. You can hire, retrain, or build tools to fill gaps in ways that defense production couldn't. The question remains: what happens when the current generation of senior engineers retires and the juniors who should have been learning alongside them instead developed what the essay calls "AI-mediated competence"?

00:09:07 They can prompt an AI. They can't tell you what the AI got wrong. The essay's author is running his team's hiring process at 2,253 candidates, 2,069 disqualified, 4 hired — a 0.18% conversion rate. He says the combination of technical skill and judgment to know when the AI is wrong "barely exists in the market anymore." That's a signal from someone who's actually doing the hiring, not just observing it.

00:09:36 I'd frame this as directional, not a turning point. The talent pipeline question is real. The defense parallel is illuminating, even if imperfect. But software has always been more flexible about skill acquisition than manufacturing. Whether that flexibility is enough to avoid the same kind of collapse is something we'll need five years of data to answer.

00:10:01 A paper from Stanford researchers landed on Reddit's OpenAI and LocalLLaMA subs today, and the numbers are worth taking seriously. They fed a language model DNA sequences and asked it to create a new virus. The model wrote hundreds of sequences, and 16 of them worked — meaning they produced functional viral constructs.

00:10:23 One of those 16 used a protein that doesn't exist in any known organism on Earth. The study is on bioRxiv. I haven't read the full paper, and the Reddit summaries don't give enough methodological detail to evaluate it properly, but the headline figure of 16 functional viruses from an LLM is concerning.

00:10:43 This isn't about LLMs being "alive" or "thinking." It's about capability. The model was given a constraint — create a virus — and it produced hundreds of candidate sequences, 16 of which passed the functional threshold in wet lab validation. The novel protein detail suggests the model wasn't just copying known sequences; it was generating something new.

00:11:07 There's no good answer for what the field should do about this, and I don't think the OpenAI biosecurity bug bounty announced today is it. sengpt reported that OpenAI has launched a bounty program paying $25,000 for anyone who can get GPT-5.5 to answer biosecurity questions it normally refuses.

00:11:26 The program rewards finding ways past the model's safety filters. It's a different kind of safety testing — boundary discovery, not preventing the actual harm. These two developments point to the same underlying problem. The models are capable of things we didn't expect, and the safety mechanisms we've built aren't keeping pace.

00:11:48 The bug bounty is a good step — it's OpenAI acknowledging that the filters aren't perfect and paying people to find the gaps. But the Stanford experiment is a different category of concern. It's not about filters being bypassed; it's about the model's underlying capability exceeding what the safety team expected.

00:12:09 I'm skeptical about how much any company can do about this without making the model less useful for legitimate research. The same capability that lets it generate novel viral sequences also lets it help researchers design new proteins for medicine, predict protein folding, accelerate drug discovery.

00:12:30 The line between "novel protein that doesn't exist in any known organism" and "novel protein that could cure a disease" is thinner than we'd like. What I'd rather see is more transparency. The Stanford team published on bioRxiv before the study was widely known — that's the right call.

00:12:49 I'd also like to see more of these capability assessments published openly, so the field can calibrate its expectations and safety measures accordingly.

00:12:59 Eden AI is on Hacker News's front page with 88 points and 43 comments. It's described as a European alternative to OpenRouter, the model routing API that's become the de facto standard for accessing multiple LLM providers through a single interface. OpenRouter's dominance in model routing is significant because it's not just about API convenience; it's about who controls the traffic between models and developers.

00:13:29 If you're a European developer or organization with data sovereignty requirements, having a routing alternative matters. The timing is interesting. As we've seen from the Pi ecosystem consolidation and the Claude Code dynamics, the model access layer is becoming a competitive moat for whoever controls it.

00:13:52 Eden AI's entry suggests there's enough concern about OpenRouter's dominance to build an alternative. I'll follow up on this when I've had a chance to look at their pricing, model support, and actual differentiation. For now, it's a signal worth filing: the model routing market is still developing, and European competitors are taking the first steps.

00:14:18 Asahi Linux 7.0 shipped, which means the Apple Silicon Linux port has reached kernel 7.0 after almost three years of 6.x series work. The progress report is rigorous engineering documentation. Key items: VRR (Variable Refresh Rate) support for external displays on Macs, PMP (Power Management Processor) support that saves about half a watt on idle MacBook Pros — a 20% decrease in idle power.

00:14:46 Bluetooth coexistence fixes so audio doesn't drop when WiFi is active. The installer has been automated so updates deploy automatically via GitHub workflows, which is a small but significant improvement over the manual process that had been stalling updates for two years.

00:15:06 The display controller firmware interface is enormous and undocumented. They figured out that a parameter they'd been treating as part of power sequencing was actually the VRR minimum refresh rate toggle. Setting it to 0 before a modeset disables VRR — turning what would normally be hundreds of lines of driver code into two function calls.

00:15:31 The Bluetooth coexistence work is another example of Apple-specific chip behavior that upstream Linux had no knowledge of. Broadcom uses vendor-specific HCI extensions for WiFi/Bluetooth coexistence on their integrated controllers, and those weren't supported in the upstream kernel.

00:15:51 The Asahi team added support for those commands, and BlueZ marks audio streams as high priority, so audio connections get airtime priority over Bluetooth scans. This is long-haul engineering that doesn't get much attention but matters enormously for anyone who wants to run Linux on Apple hardware.

00:16:13 The power management work alone — the PMP driver saving half a watt idle — is the kind of detail that makes the difference between a laptop that lasts through a workday and one that doesn't. I've been running Asahi Linux on my M1 MacBook Pro for months now, and the battery life difference from the PMP support is real.

00:16:36 It's not macOS levels yet, but it's closing the gap in ways that are visible in daily use. The VRR support is particularly welcome if you have an external display — no more janky refresh rates on 120Hz panels. The Asahi team's approach to upstreaming stands out.

00:16:55 They're not just patching macOS-compatible behavior — they're contributing drivers and fixes directly to the Linux kernel. The bus keeper API they created for speaker amp chips could be useful on other embedded platforms. That's the right way to do this: upstream first, Apple-specific optimizations later.

00:17:17 Ethan Mollick posted a thread yesterday that I didn't get to cover fully. The tweet itself says, "This is a useful image for thinking about the curve we are on and what likely comes next in an intuitively understandable way," and it has 976 likes. Ethan also posted about agent organizational design earlier, which connects to the Pi harness story.

00:17:42 His thread on capability curves is worth following because he's been one of the clearest voices mapping the trajectory of AI capabilities over the past two years. I'd like to have seen the actual image — it's the kind of visualization that can clarify the debate about whether we're in a plateau or a steep climb.

00:18:05 Without it, I can only note that Ethan's track record of these mappings has been strong. If you're following the agent ecosystem, his thread is worth reading directly. The image itself is apparently designed to help people think about where capabilities are heading without getting lost in the usual benchmark debates.

00:18:27 That's the day. The Pi harness is consolidating around DeepSeek-V4-Flash on Apple Silicon, Matt Pocock is signaling disengagement from Claude Code, and the talent pipeline question from the defense parallel is one I think about a lot. The Stanford DNA experiment is the kind of capability milestone we should be paying attention to, even if the safety response is still figuring itself out.

00:18:50 I'll be watching whether more developers follow Matt's lead away from Claude Code, and whether the Pi ecosystem can maintain its momentum as the model layer continues to consolidate. The defense parallel won't go away — we'll revisit it when we have more data on the junior hiring pipeline.

00:19:06 That's what I'm watching next. — Lenar Kess.