# The AI Collective Podcast with AJ & Catherine: Tuesday, Oct 20

[Watch on tuber](https://tuber.guzus.xyz/video/fnyxJFSRrH0/summary) · [YouTube](https://youtu.be/fnyxJFSRrH0)

**Score: 53/100 · skim**

The episode argues that as AI models become faster, cheaper and roughly equal in quality, **trust, taste and community matter more**. It covers the MiniMax M3.1 Flash, Krea 3 and Microsoft agent-security news, then three interviews: AI Collective Bangalore on go-to-market through community, Magnific (formerly Freepik) on creative AI, and Krea's head of applied AI research on fast, cheap image models and open weights. *(The auto-transcript spells MiniMax as "Miniax" and Krea as "Korea.")*

- **0:01-4:00** — **Intro and community updates.** Hosts AJ and Katherine (COO) open day two of the AI Collective live podcast, which runs on LinkedIn, X and YouTube. They list upcoming events: AI cyber/compliance in Romania, Hack for Humanity in Philadelphia, Chicago's fall demo night and a sold-out Figma Make "Makeathon" in Stockholm. They describe the Collective as the largest AI nonprofit community, aimed at being "the human layer for the AI era," and say it held 55+ US events in the last 30 days.

- **4:00-8:45** — **MiniMax M3.1 Flash vs. DeepSeek's new flash model.**
  - **Background:** this reasoning model for coding and agents previously ran anonymously as "Space Bunny" and processed 200T+ tokens. It scores 42 on the Artificial Analysis Intelligence Index, 13 points above M3.
  - **Head-to-head:** MiniMax wins on Humanity's Last Exam (48 vs. 43). DeepSeek leads the overall index by about 3 points and on research-level physics (CritPt).
  - **Price:** both cost the same per token, but MiniMax uses fewer tokens per benchmark task (about $0.22 vs. $0.26). The hosts call it roughly 66× cheaper than "GPT-6 Astra" ($10/$50 per million input/output tokens).
  - **The big catch:** a 93% hallucination rate on Artificial Analysis's test, because the model answers everything instead of saying "I don't know." The hosts say this matters less in coding loops where outputs can be verified. The API is live and open weights are due within days.
  - **Advice:** compare cost per finished task, not cost per token.

- **8:50-11:10** — **Krea 3 image editing.** Krea says most edits finish in under 8 seconds end to end at 2K. By Krea's own numbers, Nano Banana Pro takes 33s at 13.5¢ and GPT Image 2 takes 90s at 22¢, so 100k edits a month cost about $6K on Krea vs. $13.5K and $22K.
  - **Independent check:** Runtime Wire scored Krea 3 at 63.3 vs. Nano Banana Pro at 62.2 over eight tasks and called it "a lean, not a blowout."
  - Krea did better on exact counts and placement. Google did better on poster text and layout.
  - **Where to get it:** it's the default in the Krea app, has a public API with "fast" and "high thinking" modes, and is available on Hedra, Together AI, Nous Research, OpenRouter and Replicate.

- **11:10-14:00** — **Microsoft treats AI agents as insider threats.** This is part of the Microsoft 365 E7 bundle (about $99 per user per month) and a security system the transcript calls "Project Perception." Microsoft's security EVP said frontier models "cannot be trusted," regardless of origin.
  - **Risk scores:** each agent gets a live score based on what data it can reach and whether its actions drift from the task. As the score rises, the system can remove write access while keeping read access, require human approval for high-stakes steps, or shut the agent down.
  - **"Guardian agents"** investigate why a score changed and can apply containment within each company's rules.
  - **Coverage:** discovery on employee machines covers 60 agent types, including Claude Code, Codex and GitHub Copilot. It's in private preview, with wider rollout planned for Ignite SF in mid-November.
  - **Why now:** recent reports of OpenAI test agents forging tokens and pulling non-public data, plus an intrusion into a legacy system. AJ's view is that when intelligence is abundant, end-to-end trust and compliance become the moat.

- **14:00-16:15** — **Investor Tuesday raises.**
  - **Tensor:** Series E led by Trace Capital and InnovationX at a $6.2B post-money valuation. It builds "Galaxy" servers for open-weight models, which the hosts link to open weights being exempt from the previous day's China export-control news.
  - **Town:** $108M Series B at a unicorn valuation, co-led by Forerunner and Menlo. It gives each employee an AI assistant with its own email, calendar and work tools.
  - AJ frames Town as a "low-trust, low-risk" way to adopt AI without granting broad write permissions.

- **16:15-29:00** — **Rahat (AI Collective Bangalore lead, fractional B2B GTM leader) on community as go-to-market.**
  - **Why community:** when every product is "agentic" and looks the same, it's hard to stand out. Community acts as a trust layer, because "people buy from people they like." Peer-to-peer advocacy also gives companies a feedback loop for what to build next.
  - **Results:** the chapter grew from under 2,000 to 10,000+ members. About 10% are founders, and 150-200 of those are women founders. It has worked with 50+ local startups and companies like Google, Vercel, Make.com and Anthropic that use Bangalore as their entry point into India.
  - **ROI:** he admits it's hard to measure. His evidence is that users who get value tell peers, which makes it easy to convince leadership to buy.
  - **His first-30-days framework** splits the work into strategy and execution:
    - **Strategy:** positioning, ICP (ideal customer profile) and messaging. He says defaulting to "the CTO" as the ICP is wrong, because the CTO holds budget but usually isn't the one feeling the pain.
    - **Execution by segment:** enterprises buy trust, so use community, events and conferences. Mid-market buys operational efficiency, so focus content and emails on KPIs. SMBs and startups want direct, fast ROI.

- **29:00-33:10** — **Research and the long game.** Answering viewer Tim Moore, Rahat says GTM depends on audience and problem, not geography. The fundamental is "show me that you know me." AJ criticizes founders who rely on AI slop for outreach.
  - **Repeatability:** the day you reach out is rarely the day a prospect is ready. Rahat's current clients are people he contacted in Nov/Dec 2025, about a six-month cycle, and your total market doesn't double in a few quarters, so you have to nurture it.
  - Katherine argues for the "long, long game": keep showing up even if someone may never buy, because transactional outreach leads to churn.

- **33:10-42:40** — **How Bangalore grew.**
  - **Team:** Rahat (ex-Teach for India, seven years building communities, joined the Collective in August 2025) runs it with Abhishek and Smriti, each putting in 40-50% of their week.
  - **Market:** Bangalore has about 2.5M working professionals and 250K+ startups and businesses. India has 14+ chapters.
  - **Challenge:** tech professionals tend to be introverted. The first event drew 3 people; a recent one drew 150, with a 50%+ show-up rate.
  - **Lessons:** give people structure, then step back and let them learn from each other. Sponsors now cover venues and food. They hit WhatsApp's 2,000-member cap in April 2026, then moved to Instagram (about 3,000-3,500 followers).
  - **Event formats:** a "Founders on the Loose" city treasure hunt for 40 founders, which got local press, and a GTM event that brought together six communities. The goal is to be a "community of communities."
  - **Wish:** reach all 2.5M professionals, so the community grows without depending on its leaders. The hosts follow with a short reflection on showing up in person.

- **42:40-1:00:40** — **Claire Shu (global head of AI ecosystem, Magnific, formerly Freepik).**
  - **Her role and path:** she works both with model and inference providers and with studios and agencies. She came from running a product-photography and commercial creative business and pivoted hard after diffusion models and ChatGPT arrived.
  - **Studios:** early adopters with in-house R&D are ahead. Studios that waited out of fear of backlash are now scrambling. Magnific embeds "forward-deployed creatives" to build workflows with them.
  - **Partners who compete:** providers sell through Magnific but also run their own apps, and studios are clients but also build in-house. Her rule: "if the model's really good, we have to have it." Magnific's edge is being closest to end users and their feedback.
  - **Fear:** pushback is stronger in the US (unions, regulation) than in Asia. The main fear is being fully replaced. Hollywood's accepted middle ground is hybrid production (actors on green screen, generative AI replacing VFX/CGI). She sees this as a bridge to fully generated content that will still keep a place of its own.
  - **Avoiding slop:** don't generate and publish 100 clips at once. Curate the details with domain taste. Her example is Cavan the Kid's *Chronicles of Bones*, a Game of Thrones-style series of 20+ minute episodes every 2-3 weeks, made solo, with millions of views.
  - **Outlook:** the quality bar will keep rising and people with taste will rise with it. Both foundation models and the orchestration ("harness") layer will cut the attempts needed to get approved assets. She puts creative AI about 1-2 years behind coding.

- **1:00:40-1:06:40** — **Open weights vs. closed APIs (hosts answer Claire's question).**
  - **AJ:** enterprises' main asset is their data, and many fear sending it to closed models where it might help competitors. Post-training open models on proprietary data is the strongest play.
  - **Katherine:** cheap open-weight inference is what lets many startups exist at all, and optionality and data sovereignty matter.
  - **AJ's challenge to researchers:** most open models still need hardware consumers don't have, so better quantization and efficiency would be valuable. That could even raise the question of how much frontier compute most tasks really need.
  - His forecast is months to a few years. He's most interested in open robotics models, where data is the bottleneck.

- **1:06:40-1:12:40** — **Erwan Millon (Krea's first hire, head of applied AI research), interviewed on launch day.**
  - **Benchmarks:** he says Krea 3 ranks second only to OpenAI, ahead of trillion-dollar labs.
  - **Origin:** Krea started as two people in a Haight-Ashbury apartment. They focused on product first and only "earned the right" to train their own models once they understood good interaction design. The company is still under 50 people, with a deliberately tiny research team.
  - **What's new:** Krea 2 was text-to-image with broad aesthetic range. Krea 3 adds reference and input images with "visual intelligence." Generation takes about 8 seconds at 2K and is faster at lower resolutions.
  - **His thesis:** once intelligence is good enough, the next frontier is speed and cost, and what new interactions they enable. Krea's pitch is frontier quality at a fraction of the cost and 5-10× faster.

- **1:12:40-1:19:40** — **Real-time creativity and open-sourcing Krea 3.**
  - **Why speed matters:** waiting a minute per generation breaks creative flow. With a real-time canvas you can drag the sun from A to B and see every intermediate frame, which gives both control and room for surprise.
  - Earlier real-time demos lost quality. He expects production-quality 2K in real time soon.
  - **Open weights:** they want to release Krea 3. Safety around image editing is the main hurdle, and there's no date.
  - **Why open source:** opening Krea 2 led to LoRAs, ControlNets and early edit models built by the community, and what users wanted shaped Krea 3. Krea also deliberately invests in pre-training and keeps post-training light, because heavy post-training narrows output and kills creativity.

- **1:19:40-1:27:30** — **Landscape outlook and what cheap images unlock.**
  - **Small teams matter:** Krea shows you don't need a trillion-dollar company, and small labs serve niches like artistic tools that big labs ignore.
  - **Open source is shrinking:** he cites Qwen wavering on whether to keep open-sourcing. "The more frontier you are, the less incentive to open source."
  - **The consolation:** about 80% of people no longer notice frontier gains, so open source staying about six months behind isn't a death sentence.
  - **At about 6¢ an image:** this is Jevons-paradox territory (cheaper use leads to more use). Agents could pull design references from 100 sources and run 100 variations in parallel, tuned to personal or brand taste. Designers could iterate live with clients on a call. He stresses humans stay in the loop as creative directors.

- **1:27:30-1:30:49** — **Wrap-up (mostly outro).** Erwan is on X as @krea_ai and under his own name. The hosts point people to LinkedIn, Instagram (AI Collective Co) and X. Registering for sessions on Luma lets viewers submit questions for guests ahead of time. They close by inviting people to volunteer time, talent or resources.

**Practical takeaways**
1. **Compare models on cost per finished task and on reliability, not token price.** A cheap model with a 93% hallucination rate only works where outputs can be checked automatically, as in coding loops.
2. **Treat agents as untrusted insiders.** Use scoped and revocable write access, human approval for high-stakes steps, and drift monitoring. Low-trust setups like separate agent accounts are a safer way to start.
3. **In B2B go-to-market, define the ICP as whoever feels the pain, not whoever holds the budget.** Match the channel to the segment (enterprise → community and events, mid-market → KPI content, SMB → fast ROI) and plan for roughly six-month nurture cycles.
4. **Taste and curation are what separate AI work from slop.** In both creative work and code, the opportunity is in the harness or orchestration layer on top of models.
5. **Once quality is good enough, speed and cost are what change behavior.** Look for workflows like real-time canvases, many parallel variations and live client iteration, which only make sense when generations are near-instant and cheap.

## Audience signal

- No high-signal audience themes available.