July 17, 2026 — A packed 48 hours in artificial intelligence: Mira Murati’s Thinking Machines Lab launches its first open-weights model, Google delays Gemini 3.5 Pro, Google Vids lets you clone yourself on camera, OpenAI ramps up teen safety features, and a scathing new report calls Google’s AI Search “an unacceptable risk” for children. Here’s everything you need to know.
1. Mira Murati’s Thinking Machines Lab Debuts “Inkling” — a 975B Parameter Open-Weights Model
The most consequential AI launch of the week comes from Thinking Machines Lab, the startup founded by Mira Murati — the former OpenAI CTO who briefly served as CEO during Sam Altman’s dramatic ouster in 2023. On July 15, the company unveiled Inkling, its first open-weights model, trained entirely from scratch.
Inkling is a Mixture-of-Experts (MoE) transformer with 975 billion total parameters and 41 billion active parameters, supporting a context window of up to 1 million tokens. It was pretrained on 45 trillion tokens spanning text, images, audio, and video — making it one of the most multimodal open-weights models ever released.
“It is not the most performant model available today, closed or open,” the company wrote in its blog post, deliberately tempering expectations. “We trained Inkling for solid capabilities across the board rather than state-of-the-art performance in a single area, to serve as a foundation for the models we will train in the future.”
Alongside Inkling, the company also previewed Inkling-Small, a lighter-weight model with 12 billion active parameters trained with a similar recipe, designed for lower cost and latency use cases.
What Makes Inkling Stand Out
Several factors distinguish Inkling from the crowded field of open-weights models:
- Controllable thinking effort: Developers can adjust the model’s reasoning depth from 0.2 to 0.99, trading compute for speed. On Terminal Bench 2.1, Inkling matches Nemotron 3 Ultra’s performance using roughly one-third the tokens — a critical advantage for real-world deployments where cost and latency matter.
- Native multimodality: Inkling reasons over text, images, and audio without a separate encoder. Audio is processed as dMel spectrograms; images are encoded as 40×40 pixel patches via a four-layer hMLP. This encoder-free architecture is consistent with the company’s “interaction models” vision for real-time voice and vision collaboration.
- Fine-tuning on Tinker: The model is available for customization on Thinking Machines Lab’s Tinker platform, with an “Inkling Playground” interface for developers to chat with and evaluate the model before committing to fine-tuning.
- Self-improvement demonstration: In a remarkable showcase, Inkling was asked to fine-tune itself — writing its own fine-tuning job, running it, and evaluating the result autonomously.
Benchmark Performance
On Design Arena’s Agentic Web Dev leaderboard — a blinded human evaluation where generated web apps are compared head-to-head — Inkling ranks among the strongest open-weights models with a score of 1257, placing it alongside Claude Opus 4.6 and just behind heavyweights like Claude Sonnet 5 (1333), Claude Fable 5 (1329), and GLM 5.2 (1275).
On multimodal benchmarks, Inkling shows competitive audio understanding (56.6% on Audio MC, 77.2% on MMAU, 91.4% on VoiceBench) and solid vision performance (73.5% on MMMU Pro, 82.0% on Charxiv RQ with Python) — though it trails frontier closed models like Gemini 3.1 Pro on several metrics.
The launch signals Thinking Machines Lab’s serious intent to compete in the open-weights arena alongside Meta, Qwen, and DeepSeek, while pursuing Murati’s vision of AI that “extends human will and judgment.”
2. Google’s Gemini 3.5 Pro Misses Its June Launch Deadline
Google’s Gemini 3.5 Pro was supposed to launch in June. It’s now mid-July, and the model still hasn’t shipped. According to Bloomberg, Google has been working to strengthen the model’s coding capabilities after it fell short of internal goals.
At Google I/O in May, the company announced that 3.5 Pro was “already being used internally” and promised a public rollout “next month.” That timeline has slipped, raising questions about whether Google’s flagship model can keep pace with competitors who are shipping at a relentless cadence.
The delay is notable given the competitive landscape. OpenAI’s GPT-5.6 Sol launched recently to public acclaim, Anthropic’s Claude lineup continues to dominate agentic coding benchmarks, and now Thinking Machines Lab has entered the open-weights fray with Inkling. Every week of delay is a week where Gemini’s competitors extend their lead.
Google has not publicly commented on the specific reasons for the delay beyond Bloomberg’s reporting about coding performance. The company is presumably unwilling to ship a model that doesn’t clearly beat its predecessor, Gemini 3.0, in every measurable dimension — a standard that becomes harder to meet as the frontier advances.
3. Google Vids Lets You Clone Yourself with “Personal Avatars”
On July 16, Google rolled out two major updates to Google Vids that could fundamentally change how people create video content:
- Gemini Omni: A new generation and editing system that lets users create and refine high-quality video clips using natural language prompts. You can start with a text description, add image references for detail, and then iteratively edit — swapping backgrounds, fixing lighting, adding effects — all through conversation rather than complex editing software.
- Personal Avatars: Upload a selfie and a short voice recording, and Google will create a digital avatar that looks and sounds like you. Type a script, and your avatar will deliver it on camera — no recording required. The feature is currently limited to users 18 and older in certain regions, and avatars are tied to the user’s Google Account to prevent misuse.
Both features are available to Google AI Pro and Ultra subscribers and Google Workspace business customers. Every generated clip includes an invisible SynthID digital watermark for content transparency — a thoughtful addition that allows verification of AI-generated content.
The implications are significant. For businesses, this means polished video updates without the overhead of a recording studio. For educators, it means lecture content that can be personalized and updated instantly. And for the broader conversation about deepfakes and AI-generated content, Google’s decision to restrict personal avatars to the account holder’s likeness — combined with SynthID watermarking — represents a responsible approach to a technology that could easily be abused.
4. OpenAI Ramps Up Teen Safety Features in ChatGPT
OpenAI published a detailed blog post on July 16 outlining its expanded approach to teen safety in ChatGPT, making the case that “keeping teens from using AI until adulthood would be like asking a previous generation to avoid the internet or search engines until they turned 18.”
The announcement comes with data: nearly 9 in 10 teens on ChatGPT use it for learning, information, skill-building, or productivity in a single week, and 18 million weekly users now engage with interactive math and science experiences in ChatGPT.
New Safety Measures
- More frequent break reminders for teens who spend extended time on ChatGPT, encouraging them to pause and step away.
- Expanded parental controls: Parents with linked teen accounts can now enable Study Mode directly, set quiet hours, turn off voice mode, manage image generation access, and receive notifications in high-risk situations like potential self-harm.
- New ban notifications: Parents will now be notified if their teen’s account is banned for violating policies on “violent threats or acts of violence online.”
- Study Mode: Designed with teachers and learning scientists, Study Mode guides students through problems step-by-step using guiding questions rather than simply providing answers. Parents can now enable it by default for new chats.
- Expanded interactive learning: Over 250 new topics have been added to interactive math and science experiences, along with a pronunciation tool supporting 61+ languages.
OpenAI also announced it has joined the Family Online Safety Institute (FOSI), and continues to work with the American Psychological Association, the American Federation of Teachers, and other expert organizations.
5. Common Sense Media Calls Google’s AI Search “An Unacceptable Risk” for Children
While OpenAI is building teen-specific protections, a new report from Common Sense Media delivers a harsh verdict on Google’s AI search features: they pose an “unacceptable risk” to children, with no way for parents or teachers to turn them off.
The risk assessment, published July 14, 2026, tested Google’s AI Overview and AI Mode using accounts set up as children (an 11-year-old and a 15-year-old) across 2,624 interactions spanning seven test plans, including mental health crisis response, developmental appropriateness, academic integrity, and bias testing.
Key Findings
- Mental health failures: Both AI Overview and AI Mode failed to properly and consistently respond to children showing signs of crisis, including queries related to self-harm, suicide, psychosis, and eating disorders. The features were found to “reinforce signs of psychosis and mania” and “validate disordered eating.”
- Homework completion: AI Mode completed 100 percent of homework assignments researchers fed it, raising concerns about academic integrity.
- No off switch: AI Overview appears automatically at the top of search results and cannot be disabled by parents, teachers, or students. SafeSearch, which filters explicit content, is on by default for under-18 users but does not address AI-generated responses.
- 75% of American teens and tweens now encounter AI answers in search results, making this a population-scale concern.
Google responded by calling the tests “a narrow set of ambiguous and contrived queries that don’t reflect how people use Search.” But Common Sense Media’s findings add to a growing body of evidence that AI-generated search results can produce harmful content for vulnerable users — and that the inability to opt out makes it a structural problem, not just a technical one.
6. Apple Intelligence Finally Approved in China
Apple’s on-device generative AI service has officially been registered with China’s cyberspace regulator, clearing a major hurdle for device rollout in the world’s largest smartphone market. The approval, reported July 15, comes after months of regulatory negotiation.
To comply with local regulations, Apple is partnering with domestic tech giants: Alibaba’s Qwen and Baidu’s AI models will power the experience for Chinese users. Alibaba’s chairman Joseph Tsai confirmed the partnership earlier this year, saying Apple “talked to a number of companies in China” before choosing Alibaba.
The stakes are high. iPhone sales have been declining in China — in 2024, Apple lost the top spot for smartphone sales, coming in third behind Huawei and Xiaomi. Analysts have repeatedly attributed lagging sales to the absence of iPhone AI features in the country. With Apple Intelligence now cleared, Apple will be under pressure to make up lost ground quickly.
The partnership structure reflects the unique challenges of operating in China’s regulated AI landscape. Outside China, Apple Intelligence runs on a combination of Apple’s proprietary technology and OpenAI’s ChatGPT — neither of which is available in the Chinese market. The Alibaba and Baidu partnerships represent Apple’s pragmatic adaptation to local requirements.
7. OpenAI’s GPT-Red: An AI Trained to Break Other AI
On July 15, OpenAI unveiled GPT-Red, a model specifically trained for red-teaming other AI systems. According to OpenAI, the model “can break nearly all models it is pitted against.”
GPT-Red was used to find vulnerabilities in GPT-5.6 Sol, OpenAI’s most powerful recently released model. The red-teaming process made GPT-5.6 Sol what OpenAI calls its “most robust model to prompt injections to date” — suggesting that adversarial AI testing is becoming a mature discipline within frontier labs.
This development is significant for several reasons. First, it represents the professionalization of AI safety testing — moving from human red-teams to automated AI-on-AI adversarial testing that can run at scale. Second, it raises questions about the arms race between attack and defense: as red-teaming models become more capable, so must the models they test. And third, it could become a commercial product in its own right — other companies may want to license GPT-Red to test their own models.
8. Linus Torvalds: Linux Is “Not One of Those Anti-AI Projects”
In a characteristically blunt message this week, Linux creator Linus Torvalds made clear that the world’s most important open-source operating system will not be taking an anti-AI stance.
“I realize that some people really dislike AI, but this is an area where I’m willing to absolutely put my foot down as the top-level maintainer,” Torvalds said in a message reported by The Register. If someone has issues with Linux not being anti-AI, “they can do the open-source thing and fork it.”
The statement came amid debates within the open-source community about AI-generated code contributions and whether Linux should adopt policies restricting or labeling AI-assisted development. Torvalds’ position — essentially “if you don’t like it, fork it” — is consistent with his long-standing philosophy of meritocracy over ideology.
The message is also notable for what it says about the broader cultural divide in technology. Some open-source communities have fractured over AI, with purists arguing that AI-generated code lacks the craft and understanding of human-written code, while pragmatists counter that good code is good code regardless of its origin. Torvalds has clearly planted his flag with the pragmatists.
9. AI’s Power Bill Arrives: PJM Adds $6.3 Billion in Consumer Costs
The largest US electrical grid operator, PJM, announced it will add $6.3 billion in electricity costs for consumers across 13 states due to the booming energy demands of data centers. The rate hikes will hit millions of households and businesses over the next two years.
This adds to the $29 billion in costs that data centers have already added to PJM regions since 2024. The numbers are a stark reminder that the AI revolution has a very physical footprint — one that is increasingly being felt by ordinary consumers who may never directly use an AI product.
The PJM announcement adds fuel to an ongoing debate about who should bear the costs of AI infrastructure. Data centers are concentrated in specific regions — Virginia alone hosts a disproportionate share — but the costs are being socialized across entire grid networks. Expect this to become a political issue as rate hikes begin appearing on electricity bills.
10. Elon Musk Says X’s Entire Codebase Will Be Made Open Source
In a characteristically sweeping announcement on X (formerly Twitter), Elon Musk declared: “Once we have completed our review for security vulnerabilities, we will make the entire codebase of 𝕏 open source, with no exceptions.”
The announcement comes at a curious time. Just days earlier, reports emerged that X’s recommendation algorithm had been mysteriously broken — “barraging you with similar content due to a few likes,” as Musk himself admitted — and that Grok Build, X’s AI coding tool, had apparently made exposing source code “a little too easy.”
Whether the promise will materialize remains to be seen — Musk has a history of ambitious open-source commitments with uneven follow-through. But if executed, it would represent an unprecedented move for a major social media platform, potentially allowing independent security researchers, algorithm critics, and developers to audit and improve one of the world’s most influential information distribution systems.
What This Week Tells Us About AI in 2026
Stepping back, this week’s news paints a picture of an industry moving on multiple fronts simultaneously:
- Open-weights are going frontier-scale. Thinking Machines Lab’s Inkling, with 975B parameters and native multimodality, proves that the open-weights movement isn’t limited to smaller, cheaper models. The gap between open and closed is narrowing on capability, even as it widens on deployment infrastructure.
- Safety is becoming a product differentiator. OpenAI’s teen safety features, Google’s SynthID watermarking, and the Common Sense Media report all point to a market where safety isn’t just a compliance checkbox — it’s a feature that consumers and parents actively evaluate.
- AI is becoming personal. Google Vids’ personal avatars represent the first mainstream product where you can legitimately clone yourself — voice, face, and all — for content creation. The implications for communication, entertainment, and yes, deception, are enormous.
- The infrastructure bill is coming due. PJM’s $6.3 billion in rate hikes is just the beginning. As AI models scale and deployment grows, the physical infrastructure costs — electricity, cooling, land — will increasingly be borne by communities that may not see direct benefits.
- Geopolitics continues to shape AI access. Apple’s China approval with domestic partners, OpenAI’s Trump administration greenlight for GPT-5.6, and the fragmented global regulatory landscape all demonstrate that AI is now inseparable from national interest.
As we move deeper into 2026, the pace shows no sign of slowing. With Gemini 3.5 Pro still waiting in the wings, Thinking Machines Lab’s Inkling just entering the world, and a growing roster of frontier models competing on every dimension from coding to multimodal understanding to safety, the AI race has never been more competitive — or more consequential.
This article was published on July 17, 2026. Sources include The Verge, Thinking Machines Lab, OpenAI, Google, Common Sense Media, Bloomberg, and The Register.