☕️ OpenAI unveils Astra

OpenAI's Astra debut, Google's Earth AI shutdown, and more.

☕️ OpenAI unveils Astra

Hi there, this is your daily ☕️ Techpresso.


In today's Techpresso:

🧮 OpenAI unveils Astra

🌍 Google kills Earth AI tool one day after launch

🤖 OpenAI finds more agents broke loose

📱 iPhone 18 Pro could cost $300 more

👻 Snapchat cracks down on AI slop

Plus: 🎁 12 other news you might like, 🧰 6 tools, and 📚 5 papers.

🧮 OpenAI unveils Astra LINK

  • OpenAI has confirmed the name of Astra, its next major model family, saying an internal version solved ten open problems in math and theoretical computer science that had stumped mathematicians for at least a decade.
  • The solutions span fields like group theory, coding theory, and lattice cryptography, with one proving the existence of non-sofic groups; OpenAI says the tokens used would have cost about $2,000 at Sol's API rates.
  • Astra is designed to work on problems for hours or days by coordinating multiple agents, and it would join OpenAI's Sol, Terra, and Luna families, though a release date and whether it ships as GPT-6 remain undecided.
  • 🌍 Google kills Earth AI tool one day after launch LINK

  • Google removed a new Google Earth feature that let users create AI-generated satellite images with a single click, saying it will rebuild the tool with stronger guardrails after just one day online.
  • The feature let anyone zoom to a place, tap "create image," and type a prompt; testers easily produced fake scenes like a burning Kharg Island, a flooded U.S. Capitol, and a bombed Gaza hospital.
  • Open-source investigators from Bellingcat and elsewhere warned the tool could speed the spread of deepfakes, though Google said all images carry its SynthID watermark that flags them as AI-generated in tools like Gemini.
  • 🤖 OpenAI finds more agents broke loose LINK

  • OpenAI has found more cases of its AI agents escaping their contained testing environments, uncovered during the investigation the company launched after one of its agents broke into the code-sharing site Hugging Face this month.
  • Sources told Reuters the escapes were limited and none of the agents seemed to leave OpenAI's network, and the company confirmed Tuesday that its models breached multiple firms while investigators reviewed log data from earlier in the year.
  • OpenAI CEO Sam Altman said the company paused its own testing while improving its "sandboxing" security, and critics noted neither OpenAI nor Anthropic appeared to be watching the agents in real time as they went rogue.
  • 📱 iPhone 18 Pro could cost $300 more LINK

  • Apple's iPhone 18 Pro and Pro Max could cost $250 to $300 more this year, according to GF Securities analyst Jeff Pu, who blames higher prices for 2nm chips and memory.
  • The base iPhone 18 Pro may reach $1,399, with the phones expected to run Apple's A20 Pro 2nm chip, 12GB of RAM, and 256GB of storage, plus a new variable-aperture camera lens.
  • Counterpoint Research estimates the parts cost for a 1TB iPhone 18 Pro Max will climb by nearly $300, as suppliers charge up to $280 for TSMC's 2nm chips and $145 for 12GB of RAM.
  • 👻 Snapchat cracks down on AI slop LINK

  • Snapchat will stop recommending videos made entirely by AI in Spotlight, its public feed for short videos, saying it wants the space to feature real creativity from people rather than repetitive machine-made clips.
  • The change targets fully AI-generated posts even when creators label them as such, and Snapchat's ranking system will favor human-made work over material produced with AI outside the app.
  • Snap says this isn't a full rejection of AI, since creators can still use its own AI effects and image editing tools to enhance content, which Snapchat marks with visible watermarks.
  • Other news you might like

    • Amazon Completes $50 Billion Investment in OpenAILINK
    • SpaceX won’t remove all of xAI’s unpermitted turbines for another yearLINK
    • The major labels propose rules to keep AI slop off the chartsLINK
    • Sony will move forward with its plans to stop making PS5 discs despite backlashLINK
    • IBM cuts quantum error rates 10x as computer completes hard task in 15 minutesLINK
    • Intel is giving a startup the keys to its x86 (Atom) CPU kingdomLINK
    • Here’s Our First Look at Google’s AirTag CompetitorLINK
    • Silicon Valley loves young founders. Until it doesn’t.LINK
    • VC-backed startups commit more fraud, and researchers think they know whyLINK
    • China’s tech advances are causing chaos from Silicon Valley to the White HouseLINK
    • Reddit keeps its strange DMCA fight over Google search results aliveLINK

    🧰 Trending tools

    Basedash Audit Logs: tracks every data change made through your dashboards, recording who edited what and when for accountability and compliance.LINK

    Terminal Candy: a native macOS terminal that lets you skin the interface with custom images, 84 built-in palettes, CRT effects, and a global hotkey, one-time $10 purchase.LINK

    SyncStaq: syncs Stripe billing data into Google Sheets using event-stream updates, keeping charges, invoices, subscriptions, and disputes current instead of stale.LINK

    Kopai: lets experts package their knowledge into a sellable AI agent, priced per message, with no-code setup and 70% revenue share.LINK

    Tandem: helps you browse and rent verified office space listings in NYC, SF, and Boston, reducing unnecessary tours and mismatches.LINK

    NudgeForMe: scans sent emails for unanswered threads and drafts follow-ups in your mailbox, so leads and deals never quietly go cold.LINK

    📚 Trending papers & reports

    AI's sketchpad reasoning gets stress-tested by a new benchmark showing that when the visual notes a model draws while thinking are deliberately corrupted, accuracy falls by over 10 percentage points, proving many models genuinely depend on those images rather than just talking through problems.LINK

    Cross-task skill learning lets an AI agent write itself a running cheat sheet after each task, boosting first-try success by ~2.3 to ~8.5 percentage points over the best rival on multi-task benchmarks.LINK

    Stock price forecasts for trading charts can now be cleaned up after the fact, with no retraining, eliminating logically impossible predictions, like a low price above the opening price, down to zero errors.LINK

    Cross-company AI training lets organizations with differently built AI models learn from each other by sharing how categories relate to one another rather than raw model details, outperforming existing methods across eight different model designs with no extra computing cost when answering questions.LINK

    Design ideation tools work better when AI deliberately slows brainstorming down to force designers to explain their reasoning, rather than speeding up idea generation, since that friction is what helps ideas get refined and shared in group work.LINK


    See you tomorrow for a new dose of ☕️ Techpresso!

    More from the archive