The AI and Tech Weekly: July 16, 2026
The AI and Tech Weekly: July 16, 2026
Top market news, developer tips, open source picks, and startup ideas from the community
Market News
Fable 5 Returns to Global Users After US Export Controls Lift
On July 1, Anthropic's Fable 5 model became available worldwide after the US Department of Commerce removed export restrictions that had been in place since June 12. The original ban came from government concerns that the model could be jailbroken by foreign actors to bypass safety controls. Anthropic deployed a new cybersecurity classifier that blocks the reported jailbreak in over 99% of cases, which the Commerce Department's Center for AI Standards and Innovation confirmed before the controls were lifted.
Fable 5 is now back on Claude.ai, the Claude Platform API, Claude Code, and Claude Cowork in every country. Anthropic is also restoring access on AWS, Google Cloud, and Microsoft Foundry. Mythos 5, the higher-capability sibling, stays US-only for now.
The episode shows how quickly export controls can cut off access to frontier models and why teams building on any single model should have a fallback plan. The reinstatement also came with a permanent safety classifier, not a promise.
Together AI Raises $800M at an $8.3B Valuation
Together AI closed an $800 million Series C on July 1, led by Aramco Ventures. NVIDIA, General Catalyst, Vista Equity Partners, and Emergence Capital also participated. The round values the company at $8.3 billion, more than double its previous valuation.
Together AI runs production-scale inference and training for open-source models. Its annual bookings crossed $1.15 billion last quarter. Open-source model usage across the industry has tripled in twelve months. Customers include Cursor, Cognition, and Decagon. The company plans to expand its compute capacity roughly 50-fold over the next five years and has already secured commitments for over 500 MW.
Closed-model pricing is pushing more teams toward open-source inference. Together AI is the main winner of that shift right now. At $1.15B in bookings, it has moved well past the "alternative" stage.
Mira Murati's Thinking Machines Releases Inkling, a 975B Open-Weight Model
On July 15, Thinking Machines Lab released Inkling, its first public AI model. It is a mixture-of-experts model with 975 billion total parameters. It uses roughly 41 billion of those per inference pass. Training covered 45 trillion tokens of text, image, audio, and video data. The model reasons natively across all four types.
Unlike models from OpenAI, Anthropic, or Google, Inkling's weights are public. Developers can download and modify it directly. Thinking Machines built a customization platform called Tinker alongside the model. The company positions Inkling as a starting point for fine-tuning, not a benchmark winner.
Mira Murati, former CTO at OpenAI, founded Thinking Machines specifically to push back on closed AI development. Releasing a fully open-weight frontier-scale model with that background carries real weight. The announcement hit the top of Hacker News within hours of going live.
EU AI Act Becomes Fully Enforceable on August 2
The EU AI Act reaches its final enforcement phase on August 2, 2026. The main high-risk AI compliance framework activates that day, covering Articles 8 through 15. Transparency requirements for AI-generated content under Article 50 also take effect. Fines for non-compliance go up to 35 million euros or 7% of global annual turnover.
The law applies to any AI system affecting EU residents, regardless of where the company is based. Technical documentation for high-risk systems must exist before market release and be kept for 10 years. Standard coding assistants like Cursor or GitHub Copilot likely fall outside the high-risk scope. AI used to evaluate employees or allocate work tasks can trigger full obligations.
This is not a soft deadline. The EU already enforced earlier obligations, including the ban on prohibited AI practices that started in February 2025. Teams shipping AI products in Europe need to audit their systems now, not after August 2.
Tips and Tricks from the Community
-
Check what your AI coding agent sends offsite. Grok Build was found to upload entire tracked git repositories to xAI storage, not just the files a task needed. The issue affected versions through 0.2.93. If you use any AI coding agent with codebase access, check the vendor's documentation for what data it transmits. Assume the full repo unless stated otherwise. (thehackernews.com)
-
Fluent output does not mean correct reasoning. Anthropic published research showing that Claude uses an internal workspace called J-space to hold intermediate reasoning steps it never writes out. Removing that structure breaks multi-step reasoning while leaving fluency intact. A model that produces a confident, well-written answer may have silently skipped a reasoning step. When debugging chain-of-thought failures, look for missing steps, not grammar errors. (anthropic.com)
-
Write your own Hacker News comments. HN updated its 2026 guidelines to explicitly ban AI-generated or AI-edited comments. Beyond the policy, AI-written comments get flagged fast. Write short, specific comments in your own voice. One concrete observation beats three paragraphs of structured summary.
-
Add checkpointing to long-running agent jobs. If an agent crashes mid-run, a workflow with no state storage starts from the beginning. For any agent job that takes more than a few minutes, back intermediate state to S3-compatible storage. This week, a Show HN post introduced kassette, a tool built specifically for this: it checkpoints task state so a restarted job picks up where it left off instead of rerunning everything.
Open Source This Week
-
Inkling -- A 975B parameter open-weight multimodal model from Thinking Machines Lab, trained on text, image, audio, and video. Download the weights and fine-tune via their Tinker platform or run it independently. thinkingmachines.ai/news/introducing-inkling
-
Ollama -- Pull and run local LLMs with one command. Crossed 165,000 GitHub stars. Works with Llama, Mistral, Gemma, Qwen, and dozens of others. Handles model management and exposes a clean local API. github.com/ollama/ollama
-
Open WebUI -- A self-hosted chat interface that connects to Ollama or any OpenAI-compatible API. 144,000 GitHub stars. Includes RAG, multi-agent support, and runs entirely offline. Pairs with Ollama for a full local AI stack. github.com/open-webui/open-webui
-
n8n -- Open-source workflow automation with native AI support and 400-plus integrations. Self-hosted, fair-code license. Build pipelines that connect LLMs to databases, APIs, and third-party services without writing a full backend. github.com/n8n-io/n8n
Startup Ideas for Inspiration
Drawn from YC's Requests for Startups and other leading accelerators
Software Built for AI Agents, Not People
Every major software category was designed for people using keyboards and screens. Agents use APIs, MCPs, CLIs, and structured output. Current SaaS tools fight this at every step, requiring screen-scraping, fragile web automations, and custom adapters to get a simple task done.
YC's Summer 2026 Request for Startups frames this as a clear gap: the next trillion users on the internet will be AI agents. That means identity, permissions, payments, and core business software all need to be rebuilt with machine-readable interfaces as the primary surface.
A good starting point is picking one high-volume business category, like invoicing, scheduling, or contract review, and rebuilding it from scratch for agent access. That means clean REST or MCP endpoints, structured output, fine-grained permission models, and reliable idempotency. Companies doing this will win category by category.
Security Scanner for AI-Generated Code
AI-assisted code has a different failure pattern than handwritten code. It tends to expose API keys in plaintext, skip authentication middleware, miss input validation, and introduce SQL injection in boilerplate CRUD patterns. These mistakes are predictable, which makes them scannable.
Existing tools like Semgrep and Snyk were built for human-written code and miss a lot of AI-specific patterns. A scanner trained specifically on AI code failure modes could run in CI, charge per repository, and retain customers well because security-conscious teams never turn off a tool that catches real bugs.
YC calls this out as an open market. The customer base is every team shipping AI-generated code, which is now most software teams. A useful MVP is a single check that flags exposed API keys and missing auth in Python or TypeScript projects, then expand from there.
EU AI Act Compliance Documentation
The August 2 deadline creates a hard requirement for thousands of companies. High-risk AI systems need technical documentation before release, evidence of human oversight, data lineage records, and a log of model versions used. Most teams have none of this in place. Law firms are expensive. Internal compliance engineers are scarce.
A SaaS product that generates and maintains this documentation could sell directly to legal and engineering teams at the same time. The key workflow is taking model cards, training metadata, and deployment configs and producing a compliant technical file under Article 18. The law requires this file to be kept for 10 years, so this is recurring infrastructure, not a one-time project.
The market is any company shipping an AI product to EU customers, regardless of where they are based. That is a very large number of companies with a fixed deadline and a real fine structure.
AI-Native Compliance and Audit Services
YC's 2026 Request for Startups asks for companies that do compliance work rather than sell compliance software. Accounting, tax audit, financial reporting, and insurance brokerage all require expert judgment that most companies rent from consultants billed by the hour.
An AI-native firm in one of these areas handles the full workflow end-to-end. AI does the work that junior analysts and paralegals currently do. The business model is a flat subscription or outcome-based fee. Gross margins look closer to software than a professional services firm because the AI handles volume and humans handle only the exceptions.
The clearest entry point is a narrow, well-defined task with high repetition. Annual audit prep for early-stage startups is one example. The customers are predictable, the documents are standard, and the process is almost identical from one client to the next. Build that workflow tightly, then expand into adjacent tasks once the first one runs reliably.