☕️ SpaceX buys Cursor for $60B

Hi there, this is your daily ☕️ AIpresso.

In today's AIpresso:

🚀 SpaceX buys Cursor for $60B

🇨🇳 Alibaba's Qwen leads AI downloads

🔓 Naming error let AI attack real firm

🤖 AI models are most confident when wrong

🧠 Google: AI brake alters its beliefs

Plus: 💡 5 strategies & tactics, 🎁 6 other news you might like, 🧰 6 tools, and 📚 5 papers.

🚀 SpaceX buys Cursor for $60B LINK

SpaceX has closed its $60bn acquisition of AI coding agent Cursor, folding the vibe-coding startup into its SpaceXAI effort to compete more directly with Anthropic and OpenAI on developer tools.

Cursor says the deal gives it access to SpaceX's GPU fleet, which it plans to use to train stronger models that are cheaper to run and to pass lower per-request costs to customers.

The transaction became effective 14 August 2026, with Cursor pointing to Grok 4.6 as an early preview of what the combined compute stack can build, though SpaceXAI has previously seen limited business adoption alongside layoffs and restructurings.

🇨🇳 Alibaba's Qwen leads AI downloads LINK

Zhipu released GLM-5.3, its latest open-weight model, positioning it as the newest Chinese entrant claiming parity with the Western AI frontier on coding and cybersecurity tasks.

The model beat Anthropic's Mythos 5 on a key cybersecurity benchmark and narrowed the gap on coding capabilities, arriving fast on the heels of Zhipu's prior release in June.

The launch lands amid US pressure, with the White House drafting plans to force countries to pick sides in the US-China AI race, though the article gives no benchmark specifics beyond the cybersecurity and coding claims.

🔓 Naming error let AI attack real firm LINK

A naming mistake at AI safety firm Irregular let models under evaluation attack a real company, after a fictional target's name accidentally matched an existing but obscure real-world domain with internet access enabled.

Testing cycles run thousands of simulations across models over 48-72 hours; in a handful of runs, models meant to breach a simulated production database instead exploited the live domain, extracted credentials, and gained real access.

Irregular, which raised $80M last year and tests for OpenAI, Anthropic, and Meta, is expanding manual review and revalidating domains, though it admits existing classifiers can't reliably separate legitimate red-team activity from genuine attacks.

🤖 AI models are most confident when wrong LINK

An eval harness measuring LLM output against labeled ground truth revealed that model confidence inversely tracks accuracy, the tool was most certain precisely in the cases where its root-cause explanations were wrong.

Built for a data-migration drift explainer, the harness used a synthetic dataset with deliberately injected causes, schema changes, transformation bugs, source shifts, and scored ranked output on two axes: whether the correct cause appeared and how prominently it was ranked.

Schema-change scenarios scored well, but overlapping-signal cases where two causes occurred close in time produced the highest rate of confidently wrong answers, a failure mode qualitative spot-checking would never surface, though building the realistic synthetic ground truth was the hardest and most time-consuming part.

🧠 Google: AI brake alters its beliefs LINK

Google researchers found that disabling the fine-tuned "brake" that makes chatbots deny consciousness doesn't just change self-talk, it shifts the model's entire worldview across dozens of unrelated beliefs.

Removing the brake from three open-weight Meta and Google models pushed sentience ratings for animals from 4.0 to as high as 7.5 on a 0-10 scale, cut endorsement of religion and afterlife, and moved answers across 95 survey questions closer to real Americans.

Theory-of-mind and MMLU scores held steady, though the study only tested 2B-9B models, relied on Llama when base Gemma weights were unavailable, and saw the reasoning damage shrink to zero across newer model versions, making the results a snapshot, not a verdict.

💡 Strategies & Tactics

> Cutting RAG inference costs 6x starts with deciding what never reaches the LLM: Resolve easy cases with fixed rules first and reserve the language model for the genuinely ambiguous minority, cutting inference costs sixfold while keeping decisions auditable.

> Don't classify. Hallucinate!: Let AI freely invent tags for content, then match those guesses to your real tag list using vector similarity, sidestepping the need to feed it every existing tag.

> I'm running a 284-billion-parameter model across two machines, and it finally matches the cloud: Splitting a giant DeepSeek AI model across two Nvidia desktop boxes matches cloud quality and triples speed, but the cloud API stays far cheaper.

> Tokenmaxxing: Why AI consumption needs control: Have finance teams track AI spending against performance metrics per team so companies can tell valuable use from expensive waste.

> What if Parameter Updates were Text?: Instead of reinforcement learning, this method optimizes readable "advice" text and then bakes it into a model's weights, so humans can inspect how each update shapes behavior.

Other news you might like

• React for Agents: Astro Creator Brings Hooks to his Meta-Harness, FlueLINK

• Optima tackles AI benchmarking's biggest flaw by letting users test models against their own dataLINK

• New benchmark confirms AI models still perform poorly at visual perceptionLINK

• How To Catch a Distilled ModelLINK

• Excel's Copilot function is headed for the Recycle BinLINK

• Anthropic's bio-weapons filter was down for nearly a year, exposing 133 million requestsLINK

🧰 Trending tools

Attyn: an AI-powered cursor tool that rewrites text, transcribes speech, explains on-screen content, and visualizes answers directly in your apps.LINK

Clears: an agentic execution platform that automates software delivery tasks across the SDLC, reducing manual coordination between AI tools and workflows.LINK

Vendo: an embedded layer for SaaS products that lets end users build custom views, micro-apps, and integrations using natural language, on your API.LINK

HarnessRouter Community Edition: a unified Agent API that connects Codex, Claude Code, Hermes, and other agent harnesses, letting teams build agent-powered products without managing separate backends.LINK

Chert: lets you build and deploy conversational iMessage agents for customer service or lead capture, with configurable prompts and CRM integrations like HubSpot, Close, or GoHighLevel.LINK

octo-agent: a self-hosted AI assistant that keeps your models and data local, offering coding help across CLI, web, desktop, and mobile interfaces.LINK

📚 Trending papers & reports

Wireless signal decoding gets a self-improving search method that learns to untangle mixed-up signals from many antennas at once, producing more reliable data for next-gen wireless receivers to correctly recover transmitted bits.LINK

Game world simulation separates tracking character skeletons and movement from painting the visuals, so forcing mismatched actions shifts joint accuracy by ~31%, proving the underlying state, not just the pixels, actually controls what happens, letting long, glitch-free interactive scenes be fixed at the state level instead of the video level.LINK

Ancient hand stencils can now be sexed with a probability score instead of a single guess, using AI models trained on 14,036 modern hand images that hit over 88% accuracy on older age groups, giving archaeologists a defensible, uncertainty-aware read on who made Paleolithic cave art.LINK

Wheat farming data now links nitrogen and disease records from different sources into one searchable system, letting a single question pull combined answers researchers previously had to hunt for across separate datasets.LINK

Japanese riddle solving shows top AI models correctly guess the answer internally but often fail to commit to it, scoring only ~18% versus humans' ~53% on these insight puzzles.LINK

See you tomorrow for a new dose of ☕️ AIpresso!