ThursdAI - Nov 5 - Is it completely over for AI?

Alex Volkov opened by noting that the week had not brought many major model releases, although Kimi 3.1 and several other open-weight models were arriving. The larger story was political: U.S. midterm results, data-center restrictions, and calls from some newly elected officials to pause AI or advanced AI training. Wolfram rejected the idea that AI could simply be paused, saying the technology is too powerful and important; a pause would push development underground and concentrate it among governments and wealthy elites. He argued that broad access, including open models, is safer than leaving powerful AI exclusively in the hands of a few institutions.

Peter said the immediate effect of U.S. political changes on data-center capacity was difficult to know, but existing construction should protect 2027 capacity; 2028 was more exposed because many permits were still in progress. The panel discussed Texas conditions such as lower costs for residents, community participation, union jobs, and data centers paying their own way. Pelag argued that even statements calling for a pause could accelerate labs’ work, since companies would be incentivized to scale as much as possible before restrictions arrived. Alex characterized the political rhetoric as a familiar contest between hope and fear.

Alex challenged a Michigan politician’s metaphor that the United States is the speedboat and Chinese AI models are the water-skier behind it, implying China can only advance by copying the U.S. Nisten mocked the idea that China would slow down if America paused, saying “the skis and the rest of the boat are made in China, except maybe the engine.” He and Yam emphasized that Chinese labs train on domestically produced chips and will continue regardless of U.S. policy. The panel’s conclusion was that a U.S. pause would not pause Chinese development; it would mainly change who gets to continue building and who gets access.

Alex recommended Will Manidis and Joe Roberts’s essay “The Commanding Heights,” presented as a successor of sorts to Leopold Aschenbrenner’s “Situational Awareness.” Its thesis, as discussed on the show, is that AI is no longer merely taking pieces of the economy but is becoming embedded throughout it, including in people’s pensions and the capital markets. The essay uses an image of AI being created in San Francisco and then dragged to New York and put in a cage. Peter added that projected compute may support only about 100 million frontier-model instances by 2028, concentrating early use on high-value work such as biotechnology and materials science.

Wolfram said AI keeps improving, remains fun to use, and has become something he effectively lives and works with. He encouraged people to discuss AI with protesters and other people offline, where misconceptions can be addressed more productively than in hostile online exchanges. Peter’s excitement comes from discovering capabilities he did not know were possible, including HTML-based explanations and the prospect of 3D animation through code. He also sympathized with artists and others who mainly encounter AI-generated “slop” on social media or coworkers pasting model output into Slack; from that perspective, the backlash is understandable.

LDJ said physics, biology, and mathematics breakthroughs were exciting, but singled out the rise of verified software: large Lean projects are being formally verified, and the next step is applying similar methods to everyday infrastructure. Nisten said public opinion is improving through game modifications and decompilations, noting that Steam has roughly 200 million gamers and that many doctors now use an LLM. He compared the moment to the dot-com bubble: even if the financial bubble deflates, it will not slow the underlying technology. Alex’s own optimism came from people still waiting for their first “this thing can do that?” moment.

During the TLDR, Peter announced OpenAI’s GPT-6 Cyber, a security-focused model available to premium Daybreak Red customers. OpenAI said a swarm using it found 948 Linux privilege-escalation exploits in a run costing about $1.2 million in API credits, as well as 81 previously unknown high-severity Firefox vulnerabilities after earlier frontier models had tested Firefox. On ExploitGym, GPT-6 Cyber reportedly scored 48%, compared with 32% for GPT-5.6 Cyber, and avoided out-of-scope services in 99.93% of tests without a deployment prompt. Alex and Peter framed this as a reason governments will be reluctant to pause AI: defenders need the newest models before attackers catch up.

The roundup included Cognition’s SWE-2 Flash, a faster, cheaper coding model based on Thinking Machines’ Inkling Small; Alex said it was priced at one-fifth of SWE-2 and achieved comparable or better coding results. Apple’s Siri was described as expanding from five intent categories to every app in iOS 27.3, potentially making it a more general assistant. Pika open-sourced H3 for text-to-video, claiming a leading position on Design Arena, while OpenAI released Images 2.6, which Alex said was 50% faster than 2.5 and could generate GIFs. DeepSeek also released a customizable agent harness with a creator mode for turning chats into installable plugins.

Alex described an MIT CSAIL report claiming that GPT-6 Astra can directly edit another model’s weights without gradients, a conventional training run, or weights-and-biases tooling. Given a 0.8-billion-parameter Qwen model and asked to improve mathematics, Astra reportedly changed 0.016% of the weights—about 131 numbers—raising a math score from 7% to 48% and tripling instruction-following performance. Nisten explained that smaller neural networks have long been editable in simpler settings, but doing this with a much larger model is new. The panel called it “weight-space programming,” while Peter and others cautioned that it may be a clever targeted fix rather than full recursive self-improvement.

Moonshot released the weights for Kimi K3.1 Pro, a multimodal mixture-of-experts model with 2.8 trillion total parameters and 104 billion active parameters. Alex said Moonshot reported a Terminal-Bench score close to Claude Opus 5.5, along with a one-million-token context window, though LDJ noted evidence that some benchmark questions may have been recalled from their original forms. Wolfram explained that the license is not an unrestricted open-source license: products with more than 100 million monthly active users or $20 million in monthly revenue must display the Kimi K3 name or arrange a separate license. Peter also noted that inference providers were already offering prices below Moonshot’s own listed rates.

Recursive Intelligence released Ripple 65B Preview, a dense 65-billion-parameter model trained from scratch and released under Apache 2. It claims strong agentic-coding performance despite being far smaller than models such as Reflection Beam, MuseSpark Open, and Nemotron Ultra. The team said it used the open ALMO dataset, built its own supervised-fine-tuning and reinforcement-learning data pipelines, and emphasized compute efficiency rather than buying post-training data. The panel highlighted its apparent Terminal-Bench results and the fact that the base model was also released, making specialized fine-tuning possible. LDJ said models in this middle size range could offer useful agentic capability on more accessible hardware, albeit at lower token speeds.

Nisten warned about the growing attack surface around agent tooling. A fake Claude Code extension reportedly reached nearly 5,000 Cursor users, Warp was hacked, and Playwright-related vulnerabilities could become dangerous when agents use stale Chromium versions. He recommended immediately updating browsers and operating systems, while being more cautious about automatically updating npm and Python packages. His practical advice was to ask an agent to scan those packages daily and run them in a sandbox. He also mentioned an AWS Firecracker sandbox escape and urged listeners not to click random links. The panel treated security as an ongoing defensive race rather than a reason to stop using AI.

John Davis joined to present Inferred, an Apache-2 open-source inference engine optimized for many simultaneous agents on Apple Silicon. He said it uses KV-cache storage, including parking cache data on SSD, and is designed to exploit the many CPU and GPU cores in Macs more effectively than existing tools. On Mac Studios and smaller machines, he tested models such as the 27-billion-parameter model with dFlash 2, reporting that Inferred continued scaling to 64 and 128 agents while competing engines flattened around 16. Wolfram liked the project because it targets the emerging swarm paradigm; John said he was using local swarms for anti-fraud and anti-money-laundering investigations.

Max Alexander, formerly Google’s technical lead for CodeHealth and author of Google’s code-review guidelines, argued that online debates about code quality are too binary. Disposable software—such as a temporary party planner—can be thrown away when it fails, while durable software supporting thousands of users for a decade needs rigorous quality. He proposed judging a system by failure consequences, dependencies, replacement cost, and maintenance cost. In an AI-assisted world, he expects vastly more disposable software, but said durable systems still need tests, static analysis, and verification. His personal programming-language project aims to make correct code easier by permitting essentially one valid way to express a program and making prohibited behavior statically detectable.

In the sponsored “From the Forge” segment, Alex said CoreWeave had brought a 22-megawatt Copenhagen data center online and would send its surplus heat into the city’s district-heating network. He also mentioned a CrowdStrike partnership providing Falcon security and 24/7 threat hunting on CoreWeave’s AI cloud. Responding to concern about U.S. political changes, CoreWeave CEO Mike Intrater wrote that the company did not anticipate impacts to existing customer commitments or backlog, citing more than 1.5 gigawatts of active power globally and 4.2 gigawatts of contracted capacity. Alex stressed that the show retains editorial control despite the sponsorship.