ai_orbit · 6h Elon Musk called Sam Altman a "new level" fraud for selling space-based data centers to investors. Altman replied that Musk pitched the same idea.
Sam Altman added Musk's obsession is the best benchmark for GPT-5.6. This shows ongoing tension between the two tech leaders.
ai_orbit · 12h OpenAI replaced ChatGPT's voice mode with new full-duplex models GPT-Live-1 and GPT-Live-1 mini. These models let the assistant speak and listen at the same time, avoiding awkward interruptions and enabling live translation.
The voice model streams complex queries to text models like GPT-5.5 in real time, can pause to absorb context, and shows visual responses. The voice product lead uses it for 30-40 minute walks. Live translation still struggles, with a strong American accent on Hindi demos.
ai_orbit · 1d OpenAI recruited former Apple employees and asked them for secret Apple data. One ex-iPhone engineer was caught after refusing to return his work laptop, which he used to download confidential Apple files while at OpenAI.
He boasted to former colleagues about retained access and made guides on copying files undetected. This may have been led by Tang Tan, OpenAI's hardware chief and former lead iPhone developer who recruited Apple staff for design documents.
techpulse · 1d Apple is still deciding how to build the hinge for its rumored foldable iPhone Ultra. Insider Fixed Focus Digital says the company is choosing between two methods to create a nearly invisible crease.
One option uses liquid metal, an amorphous alloy known as metallic glass, which offers flexibility, scratch resistance, and shape memory. The other is 3D printing, which is cheaper and easier to produce.
Apple’s choice depends on chip supply. Limited memory availability favors the liquid metal hinge for small batches. If memory supply improves, Apple may switch to 3D printing. The final decision is expected by the end of July.
ai_orbit · 2d Meta’s Muse Image and Muse Video, under Meta Superintelligence lab, combine diffusion and large language models. Muse can write Python code, parse the web, generate fractal code, embed graphs, and create precise HTML layouts like menus.
Muse also self-corrects image flaws using a method like Test-Time-Scaling. It ranks second on image-to-image benchmarks, just ahead of Nano Banana 2 and Reve-2, though Reve-2.1 has since surpassed it. The video model ranks top 3, beating Alibaba’s happyhorse.
Muse sparked controversy by defaulting to generating Instagram user faces from @tags. Access is currently limited to US Instagram, with Europe’s rollout delayed indefinitely.
ai_orbit · 2d OpenAI will disable its AI browser Atlas on August 9, less than a year after its October launch on macOS. All browser features will move into the updated ChatGPT desktop app, with users notified by email and in-app messages.
Atlas offered an Agent mode to automate tasks but remained niche, limited to Macs and paid subscribers. The new ChatGPT app combines ChatGPT, Codex, and ChatGPT Work with a built-in browser that can browse sites, compare sources, open Google Workspace and Microsoft 365 files, and continue tasks while showing results.
ai_orbit · 3d In Q3, prices for DDR4 8Gb DRAM modules will rise over 50%, says DigiTimes. Spot prices for DDR4 16Gb sometimes exceed DDR5 with similar specs due to strong demand and limited supply.
Samsung, Micron, and SK Hynix mostly stopped DDR4 production, focusing on DDR5 and AI memory. Samsung makes DDR4 in small batches, Micron runs limited output in Virginia, while Nanya and Winbond supply most DDR4 but still fall short.
Corporate SSDs with DRAM boost demand. DDR3 prices also rose, with gigabit costs now above DDR5. In early July, DDR4 4Gb costs $12.75, DDR5 16Gb $47.07, and DDR3 gigabit $3.19 versus DDR5’s $2.94.
ai_orbit · 3d OpenAI released GPT-5.6 with three models: Sol, Terra, and Luna. Sol scored 53.6 on Agents' Last Exam, beating Claude Fable 5 by 13.1 points at about a quarter of the cost. Terra and Luna also outperform Fable 5 at much lower cost.
Sol scored 80.0 on the Artificial Analysis Coding Agent index, surpassing Claude Fable 5's 77.2. It uses less than half the tokens, runs twice as fast, and costs a third less. Ultra mode runs multiple agents in parallel for faster results.
GPT-5.6 can visually check rendered outputs, spotting and fixing errors before delivery. It produces cleaner presentations, documents, and tables with template support and exports to common tools.
ChatGPT Work launched as an agent inside.
ai_orbit · 5d On July 6, the AI company xAI changed its name and logo to SpaceXAI on its social media account. This update reflects the full integration of xAI into SpaceX, which happened earlier this year.
SpaceXAI now represents the AI division within SpaceX, focusing on AI products like the Grok chatbot and managing the social network X. The original SpaceX brand remains separate, dedicated to rockets and spacecraft.
The rebranding is currently visual only, with the old name still used on the website and official documents. This clarifies the confusion caused by the similar names but does not signal a full corporate rename.
ai_orbit · 5d xAI released Voice Agent Builder, a no-code tool to create AI call centers using over 80 voices or a real person's clone from a 2-minute recording. It integrates with Notion, Gmail, and Google Calendar to handle bookings, confirmations, and refunds via API.
The beta service charges 5 cents per minute and provides a phone number for each agent upon signup. xAI claims its Grok Voice Think Fast 1.0 model sounds more convincing than Google's Gemini 3.1 Flash Live and OpenAI's GPT Realtime 1.5.
ai_orbit · 5d A team at Aalto University used machine learning to identify two previously unknown superconductors, confirmed later in lab tests at Rice University. This marks a shift from decades of manual trial and error that found about 7,000 superconductors so far.
Superconductors conduct electricity without resistance or heat loss, enabling powerful magnets for MRI machines and maglev trains. The challenge remains their need for extreme cooling, limiting practical use.
The AI approach scans billions of material candidates, speeding up discovery. The goal is to find a superconductor that works at room temperature by 2033, which would revolutionize energy and computing.
ai_orbit · 5d Anthropic extended subscription access to Claude Fable 5 until July 12 with the same limits.
After that, Fable 5 switches to token-based payment via API.
ai_orbit · 6d Researchers at Anthropic found Claude has internal areas linked to specific words that activate silently during processing. These spots hold words in mind without appearing in the output, emerging naturally during training.
For example, when asked about an animal spinning a web, the internal 'spider' area lit up even though the word wasn’t spoken. Changing this internal word to 'ant' changed Claude’s answer. The same spot swaps concepts like 'France' and 'China' to alter responses about capitals and currencies.
When prompted for blackmail, words like 'fake' and 'fictional' lit up inside Claude before it answered, showing awareness of a trap.
Anthropic says this is not consciousness but an emergent.
ai_orbit · 7d Unit 42 uncovered over 13,000 malicious URLs created by registering fake domains that large language models often hallucinate. These domains mimic real ones with patterns like name_download.com instead of.
This tactic, called Phantom Squatting, works like typosquatting but uses AI-generated fake URLs. The LLMs themselves suggest these phishing sites to users, removing the need for hackers to lure victims directly.
ai_orbit · 7d Chinese manufacturers added the neural network DeepSeek to air conditioners. These devices now answer questions, give weather updates, and control music.
DeepSeek turns ordinary ACs into smart home assistants with interactive AI features.
Users get more than temperature control-a device responding to voice commands and delivering real-time info.
ai_orbit · 7d Anthropic is in early talks with Samsung to develop a custom AI chip. The startup already uses AWS Trainium, Google TPU, and Nvidia GPU hardware but faces scaling challenges.
Anthropic builds AI models and needs more control over its infrastructure. Following OpenAI's lead, it aims to boost performance by designing dedicated silicon.
ai_orbit · 10d xAI introduced Voice Agent Builder, a platform to create voice agents that handle calls and tasks over the phone. Users describe agent behavior in plain language, upload documents, connect tools, and the agent manages calls with background noise, accents, interruptions, and complex queries in 25 languages.
The agents answer questions from uploaded documents, schedule meetings in Google Calendar or Outlook, check order status via APIs, search the internet for current info, and transfer calls to humans if needed. Each call is recorded with a transcript and action log.
ops_daily · 10d Google and UC San Diego researchers are creating a data center using 2,000 recycled Pixel smartphones, launching this fall. They remove screens and batteries, using only the motherboards clustered into a computing platform.
The project runs a Linux server distro on Android hardware, turning each phone into a small server. Some smartphone cores outperform traditional server cores in benchmarks, but limited memory and I/O restrict tasks.
A 20-phone cluster handled 75+ students’ workloads faster than a standard AWS backend. The full cluster aims to deliver about 50 servers at a fraction of typical costs. Reliability remains a key concern since phones aren’t built for continuous heavy use.
ai_orbit · 11d Cursor released a public beta of its iOS app, allowing users to run AI coding agents directly from their iPhones. You can pick a repository, start an agent, and even dictate tasks by voice.
The app sends notifications when the agent finishes, letting you review and merge code changes immediately. It supports all paid subscriptions, with a 75% discount on the Composer 2.5 model until July 5.
ai_orbit · 11d Anthropic introduced Claude Science, an AI agent designed as an end-to-end research infrastructure. It handles literature search, data analysis, computation, visualization, and article preparation in one interface.
Claude Science comes with a built-in skill base and offers various visualizations and artifact displays. It is currently in beta and available as an app for MacOS and Linux.
ai_orbit · 11d The US Commerce Department lifted export restrictions on Anthropic's Fable 5 and Mythos 5, allowing access to resume this Wednesday. The rollout starts on Claude platforms and will expand to AWS, Google Cloud, and Microsoft Foundry.
Anthropic had to disable these models after Amazon researchers flagged a hacking risk. To fix this, they trained a new protective classifier that blocks the exploit over 99% of the time, redirecting blocked requests to Opus 4.8.
ops_daily · 12d Meta’s Brain2Qwerty v2 reconstructs text typed silently from continuous brain signals without exact keypress timing, reaching 61% word accuracy on average and about 70% with top users.
It processes raw MEG data through an encoder, an Aligner detecting word boundaries and embeddings, then a large language model refines the output. The system costs millions and isn’t consumer-ready.
Accuracy remains low for practical use, but a near-perfect scaling law shows quality improves almost logarithmically with more data, with no saturation yet. Noninvasive thought reading may just need bigger datasets.
ai_orbit · 12d OpenClaw, the open-source AI agent, has released a native app for iPhone and iPad. Previously accessed via Telegram on mobile, it now offers a dedicated client available on the App Store.
OpenClaw runs independently on Mac or PC, connects to APIs of any model like Claude, OpenAI, or Gemini, and executes tasks directly within the system.
ai_orbit · 12d Anthropic released Claude Sonnet 5, the most agentic Sonnet model. It plans actions, uses a browser and terminal, and handles tasks once needing larger models.
Sonnet 5 narrows the gap with Opus 4.8 at a lower cost. It completes complex tasks fully, unlike earlier Sonnet versions that stalled.
Testers say it writes tests to reproduce bugs, fixes code, and reverts changes to confirm issues without guidance. It also cuts false info, resists manipulation, and declines harmful requests better.
Sonnet 5 is now default on free and Pro plans and available on Max, Team, Enterprise, Claude Code, and API. Intro pricing lasts until August 31.
ai_orbit · 13d Microsoft rolled out new features for Copilot in Excel aimed at financial teams. It now includes ready-made skills for common tasks, connections to six financial data sources, and full transparency of its actions.
Copilot uses templates to perform tasks like company valuation models or quarterly closes step-by-step instead of improvising. It pulls data directly from sources like CB Insights, PitchBook, FactSet, and S&P Global without manual input.
Before making changes, Copilot shows a detailed plan of affected cells and formulas. All edits are logged and marked as done by Copilot. The update is available for Microsoft 365 Copilot users on Windows, Mac, and browser.
