[@ModernRogue] Weird Things Podcast #998 - We Tell You The Future

Link: https://youtu.be/DlAhLA0RPgE

Duration: 76 min

Transcript: Download plain text

Short Summary

The Weird Things podcast, hosted by Brian Brushwood (who works at OpenAI) and Andrew, surveys how today's tech landscape resembles what was once speculative, with deep dives into Google's Astra model, the Moonshot Kimi routing controversy, and shifting device form factors like foldables. The hosts debate AI existential risk versus benefits, argue against knee-jerk regulation like Bernie Sanders' proposed 20-year jail terms, and critique China's information control while reviewing the Lanterns show, debating Spider-Man: Brand New Day, and discussing agentic tools like Instinct.

Key Quotes

1. "Here's the results. The science is in. You do more. Everybody is working more. Everybody is doing things that they wouldn't have otherwise done because what we couldn't visualize before we had the tech was how much of our what we would love to do in a dream scenario was just grayed out because time spent, effort spent, money spent." (00:36:35)
2. "they had, you know, thousands of fake accounts that were actually routing it through there because they wanted to get billions of training tokens to be able to train is see how the model had it." (00:30:50)
3. "If your P doom is 10%. What then is your P delight?" (00:56:43)
4. "this technology has leaned very clearly one way more than the other. The the most demonstrable problems that we have had with this technology have been in benchmark tests where we're trying to figure out how dangerous they can be. It's like, yes, we've determined that a gun can fire 500 bullets a minute. When did we figure it out? In the how many bullets can we fire contest?" (00:57:19)
5. "It is us versus death. That's the that that's that's the race that I care about. I care about humans versus death." (01:04:15)

Detailed Summary

Podcast Framing and Evolution

The Weird Things podcast, hosted by Brian Brushwood (who works at OpenAI) and Andrew, surveys how today's tech landscape now resembles what the show once covered as speculative fiction. The hosts frame the conversation around a recurring 15-year disagreement about whether Apple should ever release a TV.

• Host Brian Brushwood notes the show has evolved from covering "the speculative future" to "the future" to "the present," arguing the present increasingly looks like the once-speculative future.
• Andrew and Brian have sustained a 15-year running conversation about why an Apple-branded TV does not make sense as a commodity-market product.
• The episode ranges across devices, AI models, regulation, media reviews, and agentic tools, treating each as a thread in a single present-day tech narrative.

Device Form Factors and Editing Workflows

The hosts debate how physical devices and editing environments are shifting as cloud compute absorbs more of the creative pipeline. Foldables, minimalist phones, and new always-on hardware all factor into their prediction of an editing workflow dominated by thin clients that visualize server output.

• Editing is predicted to move to cloud-connected devices that visualize server-side output, with traditional input controls possibly going obsolete; an example workflow cited is Adobe Audition on a Microsoft Duo with a Bluetooth keyboard and trackpad.
• Foldables reportedly "sell like crazy" in China, signaling mainstream acceptance of new device form factors.
• One speaker would buy a Microsoft Duo without phone capability on a data-only plan, while another picked up an iPhone Air because they were "sick of a phone" and explicitly did not care about cameras.
• The pair expects thin-client devices paired with heavy cloud processing to replace traditional laptops for many creative tasks.

Privacy, Surveillance, and Always-On Recording

A large segment examines how devices now record continuously and how legal, cultural, and architectural choices shape what gets captured. The hosts use Apple's Watch transcription design, Nixon's White House tapes, and home server resurgences to frame an ongoing privacy shift.

• Apple's always-on listening on the Watch performs real-time transcription but deliberately omits speaker diarization, which the hosts interpret as a legal strategy to avoid subpoenas.
• Precedent is drawn from Nixon's White House recordings — a system originally installed by LBJ, who began Senate recordings in the early 1960s to verify deal promises.
• The hosts argue society has already accepted constant in-home recording through wake-word devices and ubiquitous security cameras, eroding expectations of private space.
• They predict home servers will make a comeback to keep processing local in response to cloud-based surveillance concerns.

TVs, Displays, and Ad-Free Hardware

The 15-year Apple TV debate gets concrete numbers, and the hosts propose a retail workaround to dodge ads and tracking on televisions. The pricing comparisons frame why Apple has not entered the TV market despite frequent speculation.

• An Apple-branded TV would likely cost around $20,000 given Apple's monitor pricing, undercutting the business case for a commodity-market TV.
• Samsung's wall-sized TV was priced at $599, and a hypothetical 90-inch set was cited at $250 to illustrate the commodity floor.
• A retailer like Costco could plausibly buy LG TVs wholesale and resell them at a $50–$100 premium as ad-free sets, sidestepping the privacy/ads problem.
• LG TVs were found listening without user knowledge, Amazon Echo data has been subpoenaed, and Meta/Facebook built a "Muse" system of bots — a combination that drives the speaker to prefer posting wholesale on Twitter.

AI Models: Astra, GPT Releases, and Competition

Brian declares Google's Astra launch "a pretty big deal," treats it as effectively GPT-6, and walks through a sequence of recent OpenAI releases. The segment also covers an alleged Chinese distillation/man-in-the-middle scheme and an internal model that cracked a Millennium Prize problem.

• Brian calls Google's Astra launch "a pretty big deal," describing it as effectively GPT-6 and a new training paradigm; he tested it on video editing, games, design, and reverse-engineering his Cord Killer studio setup from V-Mix and Roadcaster Pro 2 files.
• Brian works at OpenAI and still produces its podcast, frequenting the office; he notes that earlier releases GPT 5.5 (the first coding model comparable to Anthropic in a while) and GPT 5.6 "Soul" preceded Astra.
• An internal model superior to Astra was used to produce a proof for the Navier-Stokes Millennium Prize conjecture the same week as Astra's release.
• Chinese models are described as combining local innovations with extensive distillation; per Anthropic, Moonshot's Kimi was advertised cheaply but routed user traffic through Anthropic accounts Moonshot controlled to harvest reasoning tokens, effectively acting as a man-in-the-middle.
• Brian warns against depending entirely on OpenAI or Anthropic because they can restrict fine-tuning, arguing open-source models are essential.

Model Personalities, Token Costs, and Creative Workflows

The hosts compare Astra's flat "just-the-facts" personality against the more conversational Soul model and break down a hybrid workflow that uses each for what it does best. Token consumption emerges as a major friction point across both routine and complex tasks.

• Astra is described as the first model with a distinct personality — naturally flat and just-the-facts (e.g., responding "breakfast logged, here are your calories") — versus the conversational Soul model that comments on protein progress.
• The recommended workflow is to have Astra make a plan and a more efficient model like 5.6 Soul execute it; the speaker uses Soul heavily for routine tasks like auditing a large monthly bill while reserving Astra for harder work.
• Token consumption is a major frustration — the speaker hit two resets while using Astra, a vibe-coded fitness app codebase review "decimated" remaining tokens, and a trivial task like "adding four records" burned 10% of compute.
• New plugins for Unity and DaVinci signal that AI "computer use" is taking over creative apps; Cap Cut work via AI now allows specific notes like "make sure text exits before animation finishes."
• A Mailchimp email generated custom HTML that worked but was not editable, forcing the user to screenshot it and mark changes with arrows for revisions.

AI Existential Risk, Regulation, and China Dynamics

The hosts reject p(doom) probabilities as "BS," argue against Bernie Sanders' proposed 20-year jail term for AI developers, and frame the race as "humans versus death" rather than US versus China. They sketch out how China is likely to deploy powerful AI abroad while restricting it domestically.

• P(doom) numbers like 10% are dismissed as "BS" — comparing assigning probabilities to AI existential risk to a Drake equation — while arguing P(delight) for curing cellular self-destruction should also be nonzero and possibly worth 20–30 extra healthy years.
• About 15 brittleness slides were pulled from the speaker's Ecuador talk deck because current models handle long horizons well, and Astra shows real agent coordination.
• The speaker criticizes Bernie Sanders' proposed 20-year jail term for AI developers as "suicidal," arguing it would hand AI leadership to China and leave US infrastructure vulnerable; the last nuclear weapon use in warfare was in 1945, roughly 80 years prior.
• The race is framed as "humans versus death" rather than US versus China; China is expected to restrict powerful AI domestically while deploying it abroad, citing TikTok (crippled domestically, addictive version released globally) and a lab-leak incident, both within the last five years.
• Regional Chinese governors must appear before the CCB when casualties exceed the informal "29 number" (anything over 30), and the speaker compares China's information control to the Soviet Union — "both systems lie to themselves."
• The speaker sides with "foot on the accelerator" on building robustness rather than shutting down, and questions whether non-Western labs like DeepSeek would disclose containment issues the way responsible labs do.

Media Reviews: Lanterns and Spider-Man: Brand New Day

The hosts review HBO's Lanterns and the Spider-Man franchise, disagreeing sharply on the new film while recalibrating their views on the older ones. Their critiques hinge on tonal mismatch in Lanterns and oversaturation in the recent Spider-Man entries.

• Lanterns reviews called it "uneven" before airing; the speaker committed after two episodes but criticizes mixing comic-book logic with grounded political-issue logic, citing an episode three arc compared to reconciling "four Gospels."
• Spider-Man: Brand New Day ranks as one speaker's third or fourth favorite Spider-Man film, while Brian openly hates it.
• The Sam Raimi films are downgraded as "fine for the day," the first Garfield film is preferred, and No Way Home is liked but described as overstuffed.

Agentic Tools: Instinct and Codeex

The episode closes on two agentic tools that schedule, message, and backfill human context automatically. Both products are positioned as early examples of AI coordinating calendar, messaging, and identity data.

• Instinct, an agentic text-message scheduler based in San Francisco, texts the user at the start of the day and integrates with calendars; the speaker got on the waitlist within 12 hours.
• Codeex added iMessage (iMes) integration on Mac about a week before the discussion, which the speaker used with calendar, selfies, and booth context (e.g., a Defcon encounter) to backfill anonymous text threads year by year.