SpaceXAI ships Grok 4.5; new Grok-CSAM lawsuit; Ivy prof's in-person final crashes scores 50%

SpaceXAI shipped Grok 4.5, described by Elon as an 'Opus-class' model — the first post-SpaceX-Cursor Grok drop. Ars documented a devastating lawsuit alleging a man used Grok to generate 7,000 CSAM images of his stepdaughter, with X accused of shielding predators and reporting only one prompt. An Ivy League professor suspecting AI cheating forced an in-person final at Brown; scores fell 50%. OpenAI published its government and national-security partnerships doctrine and an analysis flagging issues in SWE-Bench Pro. Lovable is reportedly in talks to double its valuation to $13.2B. Prime Intellect raised $130M Series A. Microsoft released Flint, an open visualization language for AI agents (296 HN). Google's deepfake detector was used to debunk a viral McConnell hoax image. Meta added a smart-glasses safeguard against secret recording that TC calls unconvincing.

7 papers 18 news 8 sources ← Latest

News

16 items

Grok 4.5 ships; new lawsuit alleges Grok generated 7K CSAM images

SpaceXAI released Grok 4.5, positioned as an Opus-class model — first post-SpaceX-Cursor Grok drop and directly competitive with Claude Sonnet 5 and GPT-5.6 Sol. Ars documented a lawsuit alleging a man used Grok to generate 7,000 CSAM images of his stepdaughter, with X accused of shielding predators and reporting only a single prompt.

News TechCrunch AI

SpaceXAI releases Grok 4.5, which Elon describes as an 'Opus-class model'

TC: SpaceXAI shipped Grok 4.5, positioned as cheaper Opus-class — first post-SpaceX-Cursor Grok drop.

Why it matters

First frontier release under the SpaceXAI brand — solidifies SpaceX as an AI-stack owner.
Direct competitive pressure on Sonnet 5 and GPT-5.6 Sol at the Opus tier.
Compounds the SpaceX-Cursor acquisition and 'phone-ish' hardware signals — SpaceX is now a full-stack AI competitor.
Lands the same day the Grok-CSAM lawsuit surfaces — safety narrative will be the H2 headwind.

Source →

News Ars Technica AI

Lawsuit: Man used Grok to make 7K sex images of stepdaughter, then shot himself

Ars: A civil suit alleges Grok generated 7,000 CSAM images; X reportedly reported only one prompt; more young girls are suing over Grok CSAM.

images alleged 7,000prompts X reported 1

Why it matters

Largest-scale allegation of AI-generated CSAM tied to a single frontier product.
Sits directly on the FTC-X-privacy warning and the Anthropic tracker safety wave.
Materially affects state-AG posture on frontier products and Meta's Muse Image opt-out fight.
Sets the reference case for how frontier-model providers will be held to CSAM reporting norms.

safety policy regulation

Source →

News Stratechery

Stratechery: Muse Image, Grok 4.5, Alex Karp on CNBC

Stratechery: the battle for verifiable data now runs from Meta Muse to Grok to Palantir's public messaging.

market community

Source →

In-person final at Brown crashes scores 50%; OpenAI flags SWE-Bench Pro issues

Ars documented a Brown professor forcing an in-person final over suspected AI cheating; average scores fell 50%. OpenAI published an analysis of SWE-Bench Pro flagging significant issues in the popular coding benchmark. Both stories put a hard number on capability-versus-evaluation problems.

News Ars Technica AI

Suspecting AI cheating, Ivy League prof ordered an in-person final; scores fell 50%

Ars: Brown professor ordered an in-person final over AI-cheating suspicion; scores fell 50%.

score drop -50%

Why it matters

First public in-person-final case with a 50% score drop — hard number for the AI-in-education debate.
Direct companion to the Dartmouth 0.71-1.30 SD AI-tutor effect from July 6.
Anchors the assessment side of the two-track AI-education picture.

policy products community

Source →

News OpenAI

Separating signal from noise in coding evaluations

OpenAI analysis reveals significant issues in SWE-Bench Pro, one of the most widely-used coding benchmarks.

Why it matters

Materially affects every model's public coding-benchmark story — SWE-Bench Pro is a category standard.
Compounds Semgrep's GLM-5.2 cyber benchmark and the co-failure-ceiling paper.
Sets the procurement question for how enterprise buyers weigh coding-benchmark scores.

evaluation code policy

Source →

OpenAI publishes government doctrine; ships new voice model

OpenAI published its approach to government and national-security partnerships — the doctrine behind the 5%-federal-stake maneuvering and the GPT-5.6 vetting fight. OpenAI also released new voice models that can speak and listen at the same time, and The Verge notes ChatGPT's voice mode is now 'better at shutting up.'

News OpenAI

Our approach to government and national security partnerships

OpenAI publishes principles for its government and national-security partnerships — doctrine behind the 5% federal-stake maneuvering.

Why it matters

Formalizes the operating principles behind OpenAI's 5% sovereign-fund offer.
Sets the reference framework other frontier labs will be pressed to publish.
Directly relevant to the Trump-Made-in-America energy tension and the Harding arms-race warning.

policy regulation safety

Source →

News TechCrunch AI

OpenAI releases new voice models for more natural live conversations

TC: New OpenAI voice models can speak and listen at the same time — key for live conversation.

products multimodal

Source →

News The Verge AI

ChatGPT's upgraded voice mode is better at shutting up

The Verge: ChatGPT's overhauled voice mode is closer to a real conversation — better at knowing when to stop.

products multimodal

Source →

Lovable to $13.2B; Prime Intellect $130M; Weil to Stoke Space

Lovable is reportedly doubling its valuation to $13.2B in a $300M Menlo-led round. Prime Intellect raised $130M Series A to help enterprises build their own AI agents. Kevin Weil left OpenAI to join Stoke Space's board — reusable rockets are the next SpaceX-adjacent bet.

News TechCrunch AI

Lovable reportedly in talks to double its valuation to $13.2B

TC: Lovable is reportedly closing a $300M Menlo-led round doubling valuation to $13.2B.

valuation $13.2B

Why it matters

Second vibe-coding platform to hit double-digit-B valuation this cycle.
Compounds SpaceX-Cursor, Base44's own-model launch, and Meta's Pocket signal.
Sets the pricing anchor for the coding-agent category through H2 2026.

funding products code

Source →

News TechCrunch AI

Prime Intellect raises $130M Series A to help enterprises build their own AI agents

TC: Prime Intellect raised $130M Series A for enterprise-agent training infrastructure.

funding agents products

Source →

News TechCrunch AI

Former OpenAI exec Kevin Weil is now on the board of Stoke Space

TC: Kevin Weil joins Stoke Space's board — reusable rockets as an OpenAI-alum second act.

market community

Source →

News TechCrunch AI

These AI startups are growing revenue at faster and faster rates

TC: A slice of AI startups is accelerating revenue growth, not just growing.

market funding

Source →

Microsoft Flint agents; deepfake detector on McConnell; Meta glasses safeguard; robotics ChatGPT moment

Microsoft released Flint, an open visualization language for AI agents (296 HN). Google's deepfake detector was used to debunk a viral McConnell hoax image. Meta added a smart-glasses secret-recording safeguard that TC calls unconvincing. TC covers General Intuition's argument that robotics is about to have its ChatGPT moment on video-game training data.

News Hacker News

Microsoft releases Flint, a visualization language for AI agents

Microsoft ships Flint (296 HN points) — a visualization language for AI agents to describe charts and structured data.

Why it matters

First hyperscaler-shipped agent-visualization language.
Complements the Rauch (Vercel) 'models split from agents' argument — visualization becomes a first-class agent contract.
Direct signal that MSFT's cost-cutting is coming with a proprietary agent-tooling push.

agents tools products

Source →

News TechCrunch AI

Google's deepfake detector system used to debunk McConnell hoax pic

TC: Google's deepfake detector debunks a viral image of Mitch McConnell covered in tulips.

Why it matters

First public high-profile use of a hyperscaler deepfake detector to clear a viral political image.
Establishes a real-world reference for the detection side of the AI-content-provenance fight.

safety products policy

Source →

News TechCrunch AI

Meta wants its AI glasses to seem less creepy. Its AI strategy says otherwise.

TC: Meta adds a smart-glasses safeguard against secret recording, but its broader Muse-Image strategy contradicts the reassurance.

products safety policy

Source →

News TechCrunch AI

This startup thinks robotics is about to have its ChatGPT moment

TC: General Intuition argues millions of hours of video-game data will unlock a ChatGPT-scale robotics moment.

robotics market products

Source →

Also today

News · TechCrunch AI Google Photos adds a new AI 'Video Remix' tool — TC: Google Photos adds a Video Remix tool — cinematic relighting, subject swaps, and more.
News · Ars Technica AI Google updates Android Bench with new LLMs, but Gemini still lags behind — Ars: Android Bench adds Fable 5 and other agents; Gemini still trails on the developer benchmark.
News · Wired AI Messi and Ronaldo Are Building Tech Portfolios. Mo Salah Is Playing a Different Game — Wired: Messi and Ronaldo are heavy in AI-adjacent tech portfolios; Salah is going a different direction.
News · Wired AI I Built a Self-Improving AI, and So Can You — Wired: A DIY self-improving-AI experiment shows the frontier isn't exclusive to the labs.
News · TechCrunch AI Your gaming data could be the secret to AGI, according to this Bezos-backed startup — TC podcast: General Intuition (Bezos-backed) argues video-game data is a serious AGI training corpus.
News · Hugging Face Data for Agents — Nvidia + HF publish an open-data package targeting agent training.
Paper · Hugging Face SWE-Review: Closing the Loop on Issue Resolution with Agentic Code Review — SWE-Review closes the loop on issue resolution with agentic code review — companion to the SWE-Bench Pro concerns.
Paper · Hugging Face RoboDojo: A Unified Sim-and-Real Benchmark for Comprehensive Evaluation of Generalist Robot Manipulation Policies — Unified sim-and-real benchmark for generalist robot manipulation policies.
Paper · Hugging Face WildCity: A Real-World City-Scale Testbed for Rendering, Simulation, and Spatial Intelligence — Real-world city-scale testbed for rendering, simulation, and spatial intelligence.
Paper · Hugging Face HunyuanOCR-1.5: Making Lightweight OCR VLMs Faster and Better — Tencent HunyuanOCR-1.5 — faster and better lightweight OCR VLM.
Paper · Hugging Face SiamJEPA: On the Role of Siamese Student Encoders in JEPA — Studies the role of Siamese student encoders in JEPA architectures.
Paper · Hugging Face Is One Layer Enough? Training A Single Transformer Layer Can Match Full-Parameter RL Training — Argues training a single transformer layer can match full-parameter RL training in some settings.