Quick Hits

• Frontier models bypass enterprise silicon: @totheagi successfully ran the full 2.8 trillion parameter Kimi K3 on 80 RTX 5090s, proving that frontier open-weight intelligence can operate on consumer GDDR7 hardware without scarce HBM.
• Training data ethics under fire: @DavidSacks called out Anthropic for hypocrisy after reports detailed their "Project Panama," which involved pirating millions of books and physically destroying purchased copies to scan them for training data.
• Agentic coding leaves the browser: @agupta highlighted Waddle Labs, a system that acts like Claude Code for physical robotics, using agents to write task-completion code in 20 minutes.
• Pricing metrics shift to value: @swyx argues that token-based pricing is obsolete, making way for cost-per-task as the only credible metric for evaluating LLM expenses.

Kimi K3 Pushes Open Weights and Consumer Hardware

The release of Kimi K3 by @Kimi_Moonshot introduces a massive 2.8 trillion parameter mixture-of-experts model with native visual understanding and a 1 million token context window. The model architecture reportedly delivers 2.5x the intelligence per unit of compute.

What truly stands out to the community is the hardware deployment. @totheagi detailed running the untuned official MXFP4 weights on 80 RTX 5090s using plain ethernet, achieving 20 tokens per second without requiring HBM. This demonstrates that the most abundant gaming GPUs on Earth can now serve frontier intelligence.

The community rapidly embraced the release. @waterloo_intern shared a deep dive into the model code, tracing the architectural lineage from GPT-2 to Kimi K3. Access is also expanding privately, with @ErikVoorhees noting that @AskVenice is hosting the model for private, zero-data-retention inference. The launch also fueled ongoing debates about open access, with @old_sound actively disagreeing with internal efforts to restrict open-weight models.

AI Training Ethics and Frontier Rumors

AI developers continue to face severe friction over training data acquisition. @DavidSacks amplified a detailed disclosure from @itsolelehmann regarding Anthropic's "Project Panama." According to the posts, Anthropic downloaded over seven million books via piracy and later purchased millions of physical books to scan. To process the volume, vendors cut off the spines and fed loose pages through industrial scanners, destroying the physical books in the process. @DavidSacks criticized the company for expecting free access to global intellectual property while complaining about competitors using its outputs.

Interestingly, physical scanning remains a requirement for niche datasets. @BradGroux noted that the SpaceXAI team is preserving rare books by scanning them the hard way, rather than cutting them up.

In the frontier research space, rumors are swirling around Ilya Sutskever's Safe Superintelligence (SSI). @nadzi_mouad pointed to claims that SSI has reached an inflection point in its research. @ZeffMax appreciated the direct communication style of the SSI team, quoting @daniellevy__ on the necessity of big computers for deep learning.

Agentic Workspaces and Enterprise GTM Transformations

Developers are moving beyond single-prompt workflows into persistent agentic environments. @fletchrichman introduced a cloud-based human and AI workspace named Type. Unlike traditional chat interfaces, agents reply by default unless the user explicitly switches to team chat. The system avoids creating an army of individual agents, opting instead for shared computers that act as spaces for memory, skills, and automations.

Autonomy is also reaching physical systems. @agupta shared a demo of Waddle Labs, an API that connects to robots and uses agents to write code for physical tasks in 20 minutes. He called it the single most impressive demo he has seen in the past year.

In the enterprise sector, this level of automation is forcing teams to consolidate operations. @chrispisarski noted that Y Combinator founders are struggling to hire go-to-market engineers who can automate entire sales workflows using Claude. The ideal candidate understands the full sales cycle and uses subagents to automate everything from outbound LinkedIn sequences to ROI model generation, bypassing expensive traditional sales tools.

Engineering Practices, Benchmarks, and Infrastructure

As models grow more capable, evaluation metrics and enterprise architectures are adapting. @swyx declared that per-token pricing died last year, urging the industry to adopt cost-per-task metrics as detailed by @ArtificialAnlys to be taken seriously.

Regarding enterprise infrastructure, @yoheinakajima highlighted a breakdown of Netflix's in-house LLM stack. Netflix relies on vLLM and NVIDIA Triton, using unified gRPC and OpenAI-compatible APIs to integrate self-hosted models into production. Perceptions of inference itself are shifting. @joshelman recommended a piece by @naveen that frames inference not merely as token generation, but as synthesized learning about human users.

In coding workflows, @MatijaSosic shared an Amazon engineer's approach of using AI to move toward not reading code at all. The key is getting system architecture, domain models, and API boundaries right upfront before involving AI. To provide context to these systems, @mattpocockuk asked developers how they are integrating an "llm-wiki" into their repositories, questioning how teams manage version mismatches between documentation and implementation.

Miscellaneous Updates

• Open source and hardware innovation: @earthtojake celebrated an open-source text-to-CAD tool crossing 10,000 stars, capable of generating STEP files and gcode. On the hardware side, @bosmeny highlighted RETINA, a vision sensor by a Duke University team that embeds neural network processing directly into camera pixels.
• Industry notes: @posthog announced a major company repositioning. Former Google executive Vic Gundotra, shared by @Scobleizer, reflected on the open-source roots of Google I/O and urged the industry to keep AI development open. Finally, @kunchenguid endorsed the new venture ArchAstro, founded by former Meta VP Calvin Grunewald.

Practical Takeaway

If you are evaluating large language models for your organization, stop using token pricing as your primary cost metric. As @swyx emphasized, cost-per-task is the only metric that reflects real-world performance. Update your benchmarks accordingly. For local or self-hosted deployments, investigate whether you actually need enterprise silicon. Testing open-weight models like Kimi K3 using MXFP4 weights on clusters of consumer GDDR7 GPUs could drastically reduce your inference infrastructure costs while maintaining frontier capabilities.