Databricks hits a $188B valuation and finds open weights match proprietary coding
Databricks reached a $188B valuation while publishing research showing open-weight models match proprietary alternatives for coding tasks. The State of Open Source AI report confirmed the broader trend, finding the open-closed capability gap has narrowed to 3.3% with open models now dominating production token volume, and Microsoft announced Project Perception, a multi-model AI security platform competing with Anthropic’s Mythos at lower cost.
Funding & Business #
Databricks Hits $188B Valuation, Extending Its Run as AI’s Favorite Second Act #
TechCrunch
Databricks raised approximately $3B from Coatue at a $188B valuation, just five months after closing a $5B Series L at $134B. The company’s pivot from big data analytics to AI enterprise platform drove the surge, bolstered by published benchmarking research demonstrating open-weight models can handle “even the highest level of task difficulty” in coding while reducing costs versus proprietary alternatives. The rapid valuation escalation – $54B in five months – reflects investor confidence in Databricks’ positioning as the enterprise bridge between open-weight models and production AI deployments.
First GPU Financiers Turn to Inference Chips in $400M Deal #
TechCrunch
General Compute secured a $400M loan from Upper90, using SambaNova’s SN50 inference chips as collateral – the first major financing deal backed by inference-specific hardware rather than GPUs. General Compute claims 16x faster inference than GPU-based clouds with lower power requirements. The deal signals growing market confidence in open-source model serving on specialized hardware and potential fragmentation of NVIDIA’s dominance as inference workloads diverge from training workloads.
Apple Files Trade Secrets Lawsuit Against OpenAI #
TechCrunch
Apple filed a trade secrets lawsuit against OpenAI alleging a pattern of misconduct reaching up to OpenAI’s chief hardware officer, with more than 400 former Apple employees now at the company. The timing is pointed: OpenAI is reportedly eyeing an IPO, and a trade secrets case from Apple creates both discovery risk and narrative liability. Whether this represents genuine IP protection or strategic positioning to complicate a rival’s public offering, the lawsuit adds a major legal overhang to the most anticipated AI IPO.
Claude Fable 5 Made Permanent in Premium Plans #
Anthropic / Simon Willison’s Weblog
Starting July 20, Anthropic will include Claude Fable 5 in all Max and Team Premium plans at 50% of rate limits, with Pro and Team Standard users retaining access via usage credits and receiving a one-time $100 credit. Fable 5 currently leads the Artificial Analysis Intelligence Index at 59.9%, ahead of GPT-5.6 Sol and Kimi K3. Transitioning the highest-performing model from limited availability to permanent inclusion compresses the premium tier – users no longer need to chase limited-access windows for frontier capability.
Open Source #
The State of Open Source AI – V1.0 #
Hacker News (452 points)
A comprehensive report finds the capability gap between open-weight and closed models has narrowed from 8.04% to 3.3%, with open models matching closed systems on coding and general tasks while trailing primarily on reasoning. Open-weight models now represent the majority of tokens on major serving platforms, with five of the highest-volume models all being open weights and Chinese-built models capturing over 45% of weekly traffic. However, adoption faces operational rather than performance barriers: only 51% of open-model teams reach production versus 63% for closed systems, primarily due to deployment complexity. The report identifies the “agentic harness” – orchestration, memory, and permission models – as the emerging battleground where closed labs maintain advantages through integrated tooling.
Security #
Microsoft Announces Project Perception for AI-Driven Vulnerability Detection #
NewsBytes / TechRepublic / Neowin
Microsoft is launching Project Perception, a multi-model AI cybersecurity platform that routes security tasks across Microsoft, OpenAI, and Anthropic models using a “model router” to optimize cost and accuracy per vulnerability type. The tool is positioned as a lower-cost alternative to Anthropic’s Mythos for enterprise vulnerability detection, led by Hayete Gallot, Microsoft’s head of security. The router pattern lets Microsoft arbitrage model pricing while maintaining a unified security product – wrapping multi-vendor AI behind Microsoft’s security brand rather than competing on individual model capability.
Regulatory & Policy #
San Francisco Orders Apple, Google to Remove AI Nudify Apps #
Ars Technica
San Francisco’s attorney general ordered Apple and Google to remove AI-powered nudify applications from their app stores, estimating the companies have collectively made millions in fees from these apps. The order shifts enforcement from targeting app developers to targeting distribution platforms, applying pressure where removal can have immediate effect. The move establishes a precedent for platform liability in AI-generated harmful content distribution that could extend to other categories of AI-generated material.
Patreon Blocks AI Scrapers with Cloudflare, Moving Beyond robots.txt #
TechCrunch
Patreon partnered with Cloudflare to actively block AI training bots from scraping creator content, reporting that weekly scraping attempts dropped from thousands to zero during testing – confirming that previous robots.txt-based requests were being systematically ignored. The platform now allows only indexing bots that direct users back to Patreon. The shift from voluntary compliance to technical enforcement marks a broader pattern: passive content protection mechanisms are being replaced by active blocking as creators and platforms recognize that polite requests have not worked.
Developer Tools #
How Smartsheet Built a Remote MCP Server on AWS #
AWS Machine Learning Blog / Smartsheet
Smartsheet’s remote MCP server uses Fargate for stateless containers, Kinesis and Flink for streaming events, and Neptune with Databricks for intelligence layers, enabling AI assistants to access Smartsheet data through natural language at enterprise scale. Since general availability, the platform reports 87% week-over-week user growth and has saved over 3B tokens through optimizations like progressive disclosure and proprietary serialization. For the MCP ecosystem, this represents the most detailed public architecture for enterprise-scale remote MCP deployment, demonstrating that the protocol can handle production workloads with proper infrastructure investment.
NeMo Automodel: Fine-Tune Diffusion Models at Scale with Diffusers #
Hugging Face Blog / NVIDIA
NVIDIA’s NeMo Automodel is a PyTorch training library for fine-tuning diffusion models directly from Hugging Face without checkpoint conversion or model rewrites, supporting models from 1.3B to 32B parameters. The tool provides memory-efficient sharding, latent caching, and parallelism configurations from single-GPU LoRA to multi-node distributed setups using FSDP2 and tensor parallelism. Making production-grade distributed training accessible for models like FLUX.1-dev and HunyuanVideo lowers the barrier for teams that need custom visual generation but lack the infrastructure engineering for distributed training.
A Scorecard for the AI Age #
OpenAI News
OpenAI CFO Sarah Friar introduces a practical framework for measuring AI ROI through four metrics: useful work, cost per successful task, dependability, and return on compute. The scorecard attempts to standardize how organizations evaluate AI deployments beyond simple token costs, addressing the gap between AI adoption enthusiasm and executives’ difficulty articulating concrete returns. For teams justifying AI infrastructure spend, having a framework from a major AI lab gives the conversation shared vocabulary – though the metrics notably favor hosted API consumption over self-hosted alternatives.
Proving the ROI of Agentic AI in Financial Services #
LangChain Blog
LangChain’s analysis of agentic AI ROI in financial services identifies that success requires layering economic intelligence atop engineering observability, tracking business-aligned KPIs like requirements extraction accuracy (95% target for RFP processing) and false positive reduction (60% for AML monitoring). The recommended stack combines LangGraph for orchestration, LangSmith for tracing, and Pay-i for KPI monitoring. For financial services teams evaluating agentic AI, the framework provides concrete benchmarks for what “good” looks like in regulated environments where vague productivity claims are insufficient.
Infrastructure #
NVIDIA Vera Rubin: Maximizing Intelligence per Dollar for Post-Training #
NVIDIA Blog
NVIDIA positions the Vera Rubin platform around “intelligence per dollar” as a metric for evaluating post-training investments, arguing that continuous model refinement for changing environments has become the primary workload in the agentic era. The company claims Vera Rubin can train the largest models with one-fourth the GPUs versus Blackwell, while Nemotron 3 Ultra achieves 71.7% on SWE-bench Verified at 550B parameters. The reframing from cost-per-token to intelligence-per-dollar aligns with NVIDIA’s incentive to sell premium hardware, but the underlying argument – that inference and post-training economics matter more than training economics for production AI – reflects genuine industry direction.
Research & Papers #
Announcing the Corrigibility Research Fund #
AI Alignment Forum
A new fund housed at Lightcone Infrastructure will award at least $200K in grants and prizes for corrigibility research in 2026, split between traditional grants (first application deadline August 23) and prizes for excellent work. Corrigibility – whether AI systems remain amenable to human correction and oversight – is increasingly relevant as agentic systems gain autonomy and tool access. The fund’s existence signals that corrigibility research remains underfunded relative to capabilities research, despite being directly relevant to the governance questions raised by production agent deployments.
Should We Benchmark Conceptual Capabilities Using Judgment Prediction Tasks? #
AI Alignment Forum
Proposes that conceptual reasoning tasks involving subjective judgments are poorly suited for direct benchmarking and should instead measure AI capabilities by predicting a specified person’s judgment on such questions. The reframing separates measuring capability (can the model predict what someone would think?) from measuring correctness (is the answer right?) for inherently subjective domains. For evaluation designers, this offers a tractable approach to benchmarking reasoning about contested questions without requiring ground truth.
Benchmarking Face Recognition without Real Faces #
Hugging Face Daily Papers
Demonstrates that synthetic face datasets can now replace real-face benchmarks for evaluating face recognition models, resolving the remaining privacy gap where even fully synthetic training pipelines still required real-face evaluation data. The paper shows synthetic evaluation benchmarks achieve comparable discriminative power to established real-face benchmarks. For teams working on face recognition in privacy-sensitive contexts, this removes the last dependency on real biometric data in the development pipeline.
Threads to Watch #
The open-weight economic argument is consolidating across stakeholders. Databricks’ $188B valuation is built on open-weight advocacy, the State of Open Source AI report shows 79% developer adoption with a narrowing capability gap, and General Compute’s $400M inference chip deal positions alternative hardware specifically for serving open models. The convergence of data platform economics, developer survey data, and infrastructure financing around the same thesis suggests open-weight viability is transitioning from debate to assumption in enterprise planning.
AI ROI measurement is maturing from narrative to framework. OpenAI’s AI Scorecard proposes standardized metrics, LangChain’s financial services analysis defines concrete KPI targets, and Smartsheet’s MCP server reports 87% weekly user growth as a quantified adoption signal. The shift from “AI is transformative” to “here is how to measure what AI actually delivered” reflects an enterprise market moving past the adoption decision and into the accountability phase.
Content enforcement against AI is shifting from passive to active. Patreon’s partnership with Cloudflare to block scrapers after robots.txt was systematically ignored, San Francisco’s order to remove nudify apps from app stores, and Apple’s trade secrets lawsuit against OpenAI all represent escalation from requests to enforcement actions – technical blocking, regulatory orders, and litigation replacing voluntary compliance and policy statements.
Sources Unavailable Today #
These sources could not be fetched today. Links point to their homepages so you can check them directly.
- Weights & Biases: Fully Connected – scrape: content not extractable