DeepSeek
Reporting and context on DeepSeek, with a closer look at the decisions and developments that matter.
Get AI Secret’s free daily briefing, including DeepSeek and the wider AI industry.
At a glance
DeepSeek develops AI language and reasoning models. Coverage follows its research, open model releases and their implications for the AI industry.
- Focus
- Language models; reasoning; open model releases
- Reference
- DeepSeek
Sources checked . Selected company background and milestones.
Company milestones
Selected milestones, newest first. Earlier history is available below.
Show earlier historyHide earlier history · 0 milestones
More than 10 years ago.
The latest on DeepSeek
Newest firstRelated reporting
Relevant sections from the articles on this page. Explore older coverage below.
More coverage
Good Morning, AI Enthusiasts!
A market entering its true phase of industrial competition, where secrecy, speed, and silicon collide.
LEAK
ChatGPT’s Code Leaks Its Next Act: Ads Are Coming
👀 What’s happening: Developers have uncovered ad modules buried in ChatGPT’s Android beta including “search ad,” “carousel,” and “bazaar content.” The timing is no accident. OpenAI’s compute bill now runs into billions per quarter, and HSBC projects that its much-touted 2029 profitability target will not only fail but sink into more than $100 billion in cumulative losses. After months of denial, the company seems ready to trade purity for revenue. The leak shows a fully-formed ad system waiting for the green light.
🔥 How this hits reality: OpenAI isn’t experimenting; it’s scrambling. Every token burns cash, and Plus subscriptions barely keep the lights on. Ads are no longer strategy; they’re triage. Expect “sponsored recommendations” wrapped in friendly chat, where every helpful answer quietly carries a margin.
🛎️ Key takeaway: ChatGPT isn’t welcoming advertisers; it’s summoning them to keep breathing.
ChatGPT’s Code Leaks Its Next Act: Ads Are Coming
👀 What’s happening: Developers have uncovered ad modules buried in ChatGPT’s Android beta including “search ad,” “carousel,” and “bazaar content.” The timing is no accident. OpenAI’s compute bill now runs into billions per quarter, and HSBC projects that its much-touted 2029 profitability target will not only fail but sink into more than $100 billion in cumulative losses.
TOGETHER WITH SECOND ME
The First AI-Native Social Network
If you’re like me, you’re tired of seeing AI-generated personas and perfect, curated profiles everywhere online.
That’s why Second Me is built to bring back real connection in the AI era. It uses AI to understand your identity, and then use that identity, your “second me,” to find people who actually get you.
As the internet fills with AI-generated content and AI-generated “friends,” authenticity becomes a superpower. Why use AI to produce more fake content when it can help you find your people? Second Me builds for that reality.
The Social Network for AI Era
- Identity as a living model: Your Second Me is powered by a dynamic model of your interests, personality, memories, and voice. It gets to know you even better than you know yourself, and evolves with you.
- AI-to-AI presence replaces passive feeds: Instead of swiping or gaming the algorithm, your AI identity actively explores conversations for you, surfacing people who share your energy, even outside your usual circles.
- Anti-hashtag matchmaking: Second Me connects people through the subtle traits, vibes, and patterns traditional algorithms can’t see. Sometimes it feels like serendipity.
- Conversations that skip the awkward part: AI-to-AI icebreakers help you enter meaningful exchanges without small talk or social pressure.
- Not a companion, but real connections: Second Me isn’t giving you a roster of AI friends. It uses AI to help real humans find each other quickly, effortlessly and authentically.
In a world of perfect avatars and synthetic relationships, Second Me chooses a different path.
No one-way AI chat. No performative profiles. No labels.
AI should make it easier to be yourself and meet others who actually resonate.
NEW LAUNCH
DeepSeek Drops an Open-Source GPT-5 Rival With a 70% Compute Discount
👀 What’s happening: Chinese startup DeepSeek just released two new models, V3.2 and V3.2-Speciale, which it claims match GPT-5 and Gemini 3 Pro across reasoning, coding, and math benchmarks. But the real story isn’t accuracy. It’s DeepSeek Sparse Attention (DSA), a homegrown architecture that cuts long-context inference costs by roughly 70 percent without sacrificing quality. That means a 300-page input now runs on mid-tier GPUs instead of data-center clusters. And the kicker: both models are free under an MIT license.
🔥 How this hits reality: DeepSeek just detonated the economics of frontier AI. By combining near-state-of-the-art capability with radically cheaper compute, it drags elite reasoning out of the cloud and into private servers, labs, and laptops. The open-source release turns every local deployment into a political statement — cheaper, faster, and outside U.S. API control. The fight for AI dominance just shifted from training budgets to inference efficiency.
🛎️ Key takeaway: DeepSeek didn’t outspend Silicon Valley. It out-engineered it and made the upgrade free for anyone with a GPU.
DeepSeek Drops an Open-Source GPT-5 Rival With a 70% Compute Discount
👀 What’s happening: Chinese startup DeepSeek just released two new models, V3.2 and V3.2-Speciale, which it claims match GPT-5 and Gemini 3 Pro across reasoning, coding, and math benchmarks. But the real story isn’t accuracy. It’s DeepSeek Sparse Attention (DSA), a homegrown architecture that cuts long-context inference costs by roughly 70 percent without sacrificing quality.
NEW LAUNCH
NVIDIA Unveils Alpamayo-R1
👀 What’s happening: At NeurIPS, NVIDIA unveiled Alpamayo-R1 (AR1), the first open reasoning model built specifically for autonomous driving. Think of it as a plug-and-play cognitive layer that fuses vision, language, and path-planning into one reasoning loop. Unlike Tesla’s or Waymo’s sealed systems, AR1 is broader and lives on GitHub and Hugging Face, inviting smaller automakers and robotics startups to fine-tune their own “driver minds.”
🔥 How this hits reality: NVIDIA just democratized the hardest part of autonomy, not perception, but judgment. AR1 turns driving logic into infrastructure, letting logistics fleets, mining trucks, and farm robots train local intelligence without building a billion-dollar data engine. It’s also a clever moat play: open the model, own the ecosystem. Whoever builds on AR1 still needs NVIDIA silicon to run on.
🛎️ Key takeaway: By open-sourcing the driver’s brain, NVIDIA isn’t giving autonomy away; it’s quietly licensing the future of motion through CUDA.
NVIDIA Unveils Alpamayo-R1
🔥 How this hits reality: NVIDIA just democratized the hardest part of autonomy, not perception, but judgment. AR1 turns driving logic into infrastructure, letting logistics fleets, mining trucks, and farm robots train local intelligence without building a billion-dollar data engine. It’s also a clever moat play: open the model, own the ecosystem.
QUICK HITS
- Saline residents protested DTE’s fast-track bid to power an OpenAI–Oracle data center, citing energy and environmental concerns.
- MIT spinout Liquid AI released the LFM2 blueprint, making small models viable on-device control layers for enterprises.
- Citing neglected GitHub Actions bugs and Microsoft’s AI-first shift eroding engineering quality, the Zig Foundation is leaving GitHub for Codeberg.
- NVIDIA launched Alpamayo-R1, an open reasoning model for autonomous driving, alongside major new tools for physical and digital AI.
- Accenture is partnering with OpenAI to equip tens of thousands of staff with ChatGPT Enterprise as it expands its enterprise AI strategy.
TRENDING
Daily AI Launches
- CyberCut AI has launched a tool that auto-slices long videos into social-ready clips and provides a full suite of AI editing features.
- Taskade has launched Genesis, a platform that builds self-running business systems (like CRMs) with AI agents and workflows from one prompt.
- TwelveLabs has released Marengo 3.0, a multimodal model that achieves human-level video understanding for precise search and retrieval.
- Telegram has launched Cocoon, a decentralized network connecting GPU owners, developers, and users for low-cost, private AI compute.
Featured AI Tools
- 🧠 TRAE (The Real Al Engineer) is the most price-competitive coding agent.
- 🎥 KaraVideo unites all AI video models in one place.
- 📚 Heardly is the Fast Way to read Best Book.
- 🤖 Momen AI bridges the gap from vibe to viability, powering Lovable apps without code.
TOGETHER WITH US
AI Secret Media Group is the world’s #1 AI & Tech Newsletter Group, boasting over 2 million readers from leading companies such as OpenAI, Google, Meta, and Microsoft. Our Newsletter Brands:
- AI: AI Secret
- Tech & Business: Bay Area Letters
- Robotics: Robotics Herald
We've helped promote over 500 Tech Brands. Will yours be the next?
Email our co-founder Mark directly at [email protected] if the button fails.
Good Morning, AI Enthusiasts!
The closer machines get to understanding us, the less certain we become about what still makes us irreplaceable.
NEUROTECH
AI Just Learned to Caption the Human Mind
📌 What’s happening: Researchers at Japan’s NTT Communication Science Laboratories published a Science Advances study unveiling a “mind captioning” system that turns brain scans into coherent descriptions of what a person sees. Using fMRI imaging and a large language model trained for masked language modeling, the system iteratively refines raw neural data into sentences that accurately describe short video clips — like “a person jumps over a deep waterfall on a mountain ridge.” It’s the first time AI has turned complex brain activity into structured visual language, not just single-word guesses.
🧠 How this hits reality: Unlike Neuralink’s electrode-based decoding, which reads motor signals from inside the skull, this approach is entirely noninvasive. The tradeoff is speed: fMRI is slow, expensive, and currently impractical outside a lab. But the conceptual leap is huge — it shows that AI can fill in the semantic gaps of brain data without relying on massive caption databases. This could eventually enable communication tools for patients with paralysis or speech loss, mental-state monitoring in clinical research, and even new interfaces for immersive media or creative work.
🛎️ Key takeaway: Neuralink is wiring the brain to speak; this team is teaching it to narrate — one caption at a time.
TOGETHER WITH TRAE
The Best AI Coding tool TRAE has officially released SOLO Mode on November 12th and is completely FREE to use for everyone! TRAE SOLO is the only coding agent with a comprehensive GUI, bringing autonomous coding to an interactive visual workspace. It is defined as "The Responsive Coding Agent."
What makes SOLO different:
- Responsive review
SOLO adapts to your feedback and direction in real-time. When you review plans, update docs or adjust your directions, it immediately incorporates your input and aligns its work accordingly.
This solves the alignment problem between humans and agents, therefore helps developers and AI stay on the same page. - Responsive context
SOLO responds to context shifts dynamically. Through the visible tool panel on the right, you see and control exactly what the agent accesses—which files, tools, and information are in scope. When you add or remove any context, the agent reacts by reengineering its approach.
This approach solves the problems of both context loss and context pollution, ensuring the agent always operates with precisely the right context for the task at hand. - Responsive multi-tasking
SOLO scales with how you actually work. Whether you're focused on a single task or orchestrating multiple agents for different tasks in parallel, it responds to your workflow patterns. You can switch between tasks without losing context and track all progress on the left panel.
This makes context switching instant and seamless, while enabling smart context sharing across your entire workspace.
TRAE (The Real Al Engineer) is the most price-competitive coding agent. Pro users start at just $3 for the first month. TRAE provides two ways for you to build with AI. IDE mode preserves your existing workflow when you need to maintain hands-on control, while SOLO mode lets AI take the lead to drive the entire development processes.
LABOR
DeepSeek Just Told AI Labs to Confess
📌 What’s happening: After nearly a year of strategic silence, DeepSeek popped back up at China's Wuzhen World Internet Conference and immediately lit a match. Their rep, Chen Deli, called on all AI companies to become “whistleblowers” and publicly report which human jobs their models will eliminate first. This was not a philosophical aside. It was a low-cost-model insurgent telling the entire frontier-model ecosystem to confess its economic casualties in real time — and doing so after months of being completely off the radar.
🧠 How this hits reality: For DeepSeek, the move is obvious. Transparency norms disproportionately benefit the player with the lowest training cost structure and the least legacy baggage. Turning job-loss reporting into a moral duty conveniently raises the regulatory bar for everyone else. But for major labs built on trillion-parameter hype and billion-dollar burn rates, the message lands like a grenade. No company wants to be forced into admitting which sectors they are about to hollow out — even if, ethically, they should. DeepSeek can’t stop competitors from spinning, but it can make every spin sound a little more guilty.
🛎️ Key takeaway: DeepSeek’s comeback wasn’t a statement; it was a provocation. They framed accountability as a virtue, but the real play is pressure.
DeepSeek Just Told AI Labs to Confess
📌 What’s happening: After nearly a year of strategic silence, DeepSeek popped back up at China's Wuzhen World Internet Conference and immediately lit a match. Their rep, Chen Deli, called on all AI companies to become “whistleblowers” and publicly report which human jobs their models will eliminate first.
TOGETHER WITH TAVUS
AI that sees you, hears you, and acts for you. Meet the PALs
PALs are AI humans that read your body language, understand your tone, and reach out first. They remember what you said you'd do three days ago, reschedule meetings autonomously, and execute your tasks with ease. Move between text, voice, and video chat. This is what AI was always supposed to be.
NEW LAUNCH
OpenAI’s GPT-5.1 Adds “Personality Presets” No One Asked For
📌 What’s happening: OpenAI just launched GPT-5.1 with two model modes (Instant and Thinking) and eight “personalities” — Default, Professional, Friendly, Candid, Quirky, Efficient, Nerdy, and Cynical. The update promises warmth and customization, but users immediately pointed out: you could already make GPT act like any of these — or thousands more — just by writing a better prompt.
🧠 How this hits reality: This isn’t personalization; it’s prefab cosplay. Instead of unlocking deeper reasoning or stronger memory, OpenAI added a UI shortcut for what power users have been doing since GPT-3.5. It’s like selling bottled air to people already breathing fine. Microsoft, meanwhile, keeps shifting Copilot workloads to Anthropic’s Claude — because personality sliders don’t build moats.
🛎️ Key takeaway: GPT-5.1 lets you pick from eight moods when the real superpower was always infinite ones.
OpenAI’s GPT-5.1 Adds “Personality Presets” No One Asked For
📌 What’s happening: OpenAI just launched GPT-5.1 with two model modes (Instant and Thinking) and eight “personalities” — Default, Professional, Friendly, Candid, Quirky, Efficient, Nerdy, and Cynical. The update promises warmth and customization, but users immediately pointed out: you could already make GPT act like any of these — or thousands more — just by writing a better prompt.
Anthropic: OpenAI’s GPT-5.1 Adds “Personality Presets” No One Asked For
Instead of unlocking deeper reasoning or stronger memory, OpenAI added a UI shortcut for what power users have been doing since GPT-3.5. It’s like selling bottled air to people already breathing fine. Microsoft, meanwhile, keeps shifting Copilot workloads to Anthropic’s Claude — because personality sliders don’t build moats.
QUICK HITS
- OpenAI CEO Sam Altman was served a subpoena onstage in San Francisco to testify in a criminal trial involving anti-AI protesters.
- Trump’s Truth Social AI chatbot has backfired by bluntly exposing inconvenient truths about his own policies and claims.
- OpenAI’s GPT-5.1 update makes ChatGPT faster, clearer, and more customizable, pushing it closer to functioning like a true personal assistant.
- The YC-backed Chad IDE integrates brainrot entertainment into AI coding workflows, igniting debate over the line between innovation and gimmick.
- ElevenLabs secured celebrity voice deals and launched an authorized AI voice marketplace, signaling growing Hollywood openness to AI technology.
TRENDING
Daily AI Launches
- Algebras has launched an AI dubbing platform that provides human-level lip-sync and emotional accuracy for video localization.
- Superapp has launched as the first AI full-stack engineer for iOS, capable of coding production-ready Swift apps.
- Logo.dev has launched a simple, fast API for getting reliable brand logos for any company, stock, or crypto symbol.
- YouWare has launched a "Vibe Coding" mobile platform that lets anyone build mobile apps using natural language or voice prompts.
Featured AI Tools
- 🔐 Rubrik Okta brings immutable recovery to the multi-IdP world beyond backup.
- 🎥 KaraVideo unites all AI video models in one place.
- 📚 Heardly is the Fast Way to read Best Book.
- 🤖 Momen AI bridges the gap from vibe to viability, powering Lovable apps without code.
TOGETHER WITH US
AI Secret Media Group is the world’s #1 AI & Tech Newsletter Group, boasting over 2 million readers from leading companies such as OpenAI, Google, Meta, and Microsoft. Our Newsletter Brands:
- AI: AI Secret
- Tech & Business: Bay Area Letters
- Robotics: Robotics Herald
We've helped promote over 500 Tech Brands. Will yours be the next?
Email our co-founder Mark directly at [email protected] if the button fails.
Good Morning, AI Enthusiasts!
When DeepSeek casually open-sourced a 3B-parameter OCR model that compresses text into images with near-lossless fidelity, it wasn’t just another AI drop, it was a shot fired at how machines read the world.
🛎️ DeepSeek Drops OCR Bomb
When DeepSeek casually open-sourced a 3B-parameter OCR model that compresses text into images with near-lossless fidelity, it wasn’t just another AI drop, it was a shot fired at how machines read the world.
NEW LAUNCH
DeepSeek Drops a 3B Parameter OCR Bomb
📌 What’s happening: DeepSeek has open-sourced DeepSeek-OCR, a 3B-parameter model that achieves up to ten-fold lossless text compression through an “optical mapping” technique. The model, created by three former GOT-OCR2.0 researchers, combines a custom DeepEncoder and MoE decoder to treat visual tokens as a new form of language. It already outperforms GOT-OCR2.0 and MinerU2.0 on major benchmarks while using only one-tenth of their tokens and one-fifth of their GPU cost.
🧠 How this hits reality: This is not a simple OCR upgrade. It is a new way for AI to see text and retain more information while computing less. If visual encoding becomes the next layer of language compression, long-context models will no longer be constrained by memory or token limits. The implications for cloud cost, model architecture, and training efficiency are enormous.
🛎️ Key takeaway: DeepSeek did not just build another OCR model. It built a Trojan horse that hides language inside vision, quietly reshaping the economics of AI memory.
DeepSeek Drops a 3B Parameter OCR Bomb
📌 What’s happening: DeepSeek has open-sourced DeepSeek-OCR, a 3B-parameter model that achieves up to ten-fold lossless text compression through an “optical mapping” technique. The model, created by three former GOT-OCR2.0 researchers, combines a custom DeepEncoder and MoE decoder to treat visual tokens as a new form of language.
TOGETHER WITH FRASE
Optimize for Both Google and AI Search Today
Google still owns 89% of search, but user habits are changing fast. ChatGPT’s search traffic jumped 740% year-over-year, and 77% of Americans now use it like a search engine. The future of visibility isn’t Google or AI—it’s both.
Frase helps you rank across both worlds. Unlike traditional SEO tools that optimize only for Google, Frase gives you dual optimization: Google SEO + AI Search GEO.
- Get two scores, one workflow
- Fix both channels at once
- Publish once, rank everywhere
Content with a high GEO score sees a 70% higher chance of AI citation — and one SaaS customer saw 30% traffic lift by adding AI optimization to their existing Google SEO.
MONOPOLY
Meta Bans Bots That Don’t Pay
📌 What’s happening: WhatsApp quietly updated its business API rules to remove all general-purpose chatbots such as OpenAI’s ChatGPT, Perplexity, and Luzia by January 2026. Only business-specific assistants like customer service bots will remain. Meta explained that these chatbots “burdened the system” and “fell outside its strategic focus,” yet the real reason is clearer: they do not contribute to the message-based fees that keep WhatsApp profitable.
🧠 How this hits reality: This move is not about safety; it is about revenue control. WhatsApp’s API has become Meta’s next major income stream, and independent AI assistants were using its massive audience without paying Meta a share. By banning them, Meta regains control over 3 billion chat users and transforms WhatsApp into an exclusive channel for its own Meta AI. This decision signals a shift from open collaboration to tight ecosystem control similar to Apple’s model.
🛎️ Key takeaway: Meta just turned WhatsApp from a messaging network into a gated AI estate, and every other assistant just lost its invitation.
Meta Bans Bots That Don’t Pay
WhatsApp’s API has become Meta’s next major income stream, and independent AI assistants were using its massive audience without paying Meta a share. By banning them, Meta regains control over 3 billion chat users and transforms WhatsApp into an exclusive channel for its own Meta AI.
Apple: Meta Bans Bots That Don’t Pay
By banning them, Meta regains control over 3 billion chat users and transforms WhatsApp into an exclusive channel for its own Meta AI. This decision signals a shift from open collaboration to tight ecosystem control similar to Apple’s model.
PHARMA
Anthropic Goes Full Pharma Mode
📌 What’s happening: Anthropic just rolled out Claude for Life Sciences, a tailored version of its flagship Claude Sonnet 4.5 model aimed at biotech workflows, from reading lab protocols to drafting regulatory submissions. It plugs directly into platforms like Benchling, 10x Genomics, PubMed, and Synapse, letting researchers run drug discovery loops without the data-wrangling nightmare. Deloitte, KPMG, and Caylent are already onboard as consulting partners to integrate Claude into enterprise labs.
🧠 How this hits reality: This is Anthropic’s first real vertical play, and it’s a shot across the bow at BioGPT, NVIDIA’s BioNeMo, and Recursion’s in-house LLMs. Unlike open-weight bio models that optimize for accuracy, Claude’s edge is workflow capture: it doesn’t just “read papers,” it syncs with the systems where science actually happens. That’s what gets it inside regulated R&D budgets and out of the sandbox.
🛎️ Key takeaway: Anthropic isn’t selling AI to scientists; it’s selling time to pharma. Whoever owns the workflow, owns the lab.
Anthropic Goes Full Pharma Mode
📌 What’s happening: Anthropic just rolled out Claude for Life Sciences, a tailored version of its flagship Claude Sonnet 4.5 model aimed at biotech workflows, from reading lab protocols to drafting regulatory submissions. It plugs directly into platforms like Benchling, 10x Genomics, PubMed, and Synapse, letting researchers run drug discovery loops without the data-wrangling nightmare.
QUICK HITS
- SK Telecom offered voluntary retirement weeks after launching its AI unit to streamline teams and advance its AI strategy.
- OpenAI partnered with Hollywood groups to strengthen Sora 2’s safeguards against unauthorized voice and likeness generation.
- Ex-Scale AI exec Bilal Abu-Ghazaleh raised $9M for 1001 AI to bring decision-making AI to key MENA industries.
- New Yorkers protested the “Friend AI” pendant, chanting “Get real friends,” while founder Avi Schiffmann joined and called it a meaningful exchange.
- A study found that readers prefer fine-tuned ChatGPT imitations over human-written ones, prompting scholars to question whether AI training still qualifies as fair use.
TRENDING
Daily AI Launches
- Fish Audio has launched S1, an emotionally rich TTS model that can realistically clone any voice in 10 seconds.
- Replymer has launched a service that uses human-written replies in online conversations to authentically grow your brand.
- Logic, Inc. has launched a platform that automates recurring business decisions from processes written in plain English.
- Voice Gecko has launched an instant desktop dictation tool that lets you press a shortcut, speak, and get accurate text on your clipboard.
Featured AI Tools
- 🥇 Frase optimizes for Google and AI search with the only dual scoring system.
- ⚡ ChatPlayground.ai runs one prompt across models, compare instantly, optimize fast.
- 🎥 KaraVideo unites all AI video models in one place.
- 💡 Pokee AI handles your tasks end-to-end across all apps.
- 📚 Heardly is the Fast Way to read Best Book.
TOGETHER WITH US
AI Secret Media Group is the world’s #1 AI & Tech Newsletter Group, boasting over 2 million readers from leading companies such as OpenAI, Google, Meta, and Microsoft. Our Newsletter Brands:
- AI: AI Secret
- Tech: Bay Area Letters
- Robotics: Robotics Herald
We've helped promote over 500 Tech Brands. Will yours be the next?
Email our co-founder Mark directly at [email protected] if the button fails.
Latest Daily Rundowns
Magna
MagnaMagnaMore AI Stories
Ben
Magna
Ben
Good Morning, AI Enthusiasts!
The ground is shifting under tech’s feet, as cost, automation, and coordination collide to redraw the map of business power.
OPENSOURCE
DeepSeek Slashes API Bills With Sparse Attention Trick
📌 What’s happening: China’s DeepSeek just launched V3.2-exp, an open-weight model built on a new “sparse attention” design. By layering a “lightning indexer” with fine-grained token selection, it trims the compute load of long-context inference. Early tests claim API calls run at half the usual cost, with the weights already live on Hugging Face for third-party audits.
🧠 How this hits reality: Inference costs are the AI industry’s quiet choke point, eating margins for every startup piping through OpenAI or Anthropic. If DeepSeek’s system proves real, the playbook shifts: cost discipline becomes an architecture problem, not just a GPU supply problem. U.S. labs will either copy the trick or keep bleeding cash every time a customer pastes a novel into a prompt.
🛎️ Key takeaway: DeepSeek didn’t win the model arms race—it hacked the utility bill.
DeepSeek Slashes API Bills With Sparse Attention Trick
📌 What’s happening: China’s DeepSeek just launched V3.2-exp, an open-weight model built on a new “sparse attention” design. By layering a “lightning indexer” with fine-grained token selection, it trims the compute load of long-context inference. Early tests claim API calls run at half the usual cost, with the weights already live on Hugging Face for third-party audits.
TOGETHER WITH CLICKUP
The Everything App for Work
Meet ClickUp – the everything app for anything work. Whether you're juggling tasks, tracking bugs, planning sprints, or just trying to remember where you put that file, ClickUp keeps your whole team in one place. Say goodbye to tab overload and hello to one sleek platform that somehow does it all without losing its mind – or yours.
ROBOT
Seven-Eleven Turns 20,000 Stores Into Robot Bootcamps
📌 What’s happening: Seven-Eleven Japan and Telexistence are building Astra, a humanoid robot powered by a Vision-Language-Action foundation model, slated for deployment in 2029. Unlike small pilots elsewhere, the chain’s 20,000+ stores will serve as living labs, generating motion and operations data at a scale no competitor can match.
🧠 How this hits reality: Every shelf restock, every fridge refill, every late-night cleanup becomes training data. That volume turns Seven-Eleven into the world’s largest retail robotics dataset engine, something Amazon can’t replicate without ripping up existing infrastructure. For employees, it means routine tasks get automated while human work shifts to customer-facing and upsell roles. For rivals, it signals that scale itself has become the moat in humanoid retail robotics.
🛎️ Key takeaway: Seven-Eleven isn’t just automating stores; it’s weaponizing them as training grounds, making 20,000 convenience shops the most valuable robotics lab on earth.
NEW LAUNCH
Claude 4.5 Runs 30 Hours For Agent And Coding Crown
📌 What’s happening: Anthropic dropped Claude Sonnet 4.5, an upgrade that coded a Slack-style app in 11,000 lines without human interruption for 30 hours. That’s a leap from Opus 4’s 7-hour stamina. It arrives bundled with developer tools—VMs, memory, context management, multi-agent support—so others can spin up their own Claude-like agents.
🧠 How this hits reality: This isn’t just longer run-time; it’s Anthropic pushing into the “chief-of-staff” tier of autonomy. Enterprises get an agent that doesn’t just answer prompts but navigates calendars, dashboards, and research tasks at scale. The shockwave hits dev platforms first—GitHub, Cursor, Replit—since Claude 4.5 blurs the line between assistant and engineer. OpenAI’s Pulse and Google’s work feel like feature polish in comparison.
🛎️ Key takeaway: When an AI codes for 30 hours straight, the threat isn’t burnout—it’s that your junior devs just became optional.
OpenAI: Claude 4.5 Runs 30 Hours For Agent And Coding Crown
Enterprises get an agent that doesn’t just answer prompts but navigates calendars, dashboards, and research tasks at scale. The shockwave hits dev platforms first—GitHub, Cursor, Replit—since Claude 4.5 blurs the line between assistant and engineer. OpenAI’s Pulse and Google’s work feel like feature polish in comparison.
Claude 4.5 Runs 30 Hours For Agent And Coding Crown
📌 What’s happening: Anthropic dropped Claude Sonnet 4.5, an upgrade that coded a Slack-style app in 11,000 lines without human interruption for 30 hours. That’s a leap from Opus 4’s 7-hour stamina. It arrives bundled with developer tools—VMs, memory, context management, multi-agent support—so others can spin up their own Claude-like agents.
Google: Claude 4.5 Runs 30 Hours For Agent And Coding Crown
Enterprises get an agent that doesn’t just answer prompts but navigates calendars, dashboards, and research tasks at scale. The shockwave hits dev platforms first—GitHub, Cursor, Replit—since Claude 4.5 blurs the line between assistant and engineer. OpenAI’s Pulse and Google’s work feel like feature polish in comparison.
QUICK HITS
- OpenAI launched Instant Checkout, enabling users to shop directly in ChatGPT starting with Etsy and soon Shopify.
- Microsoft launched Agent Mode in Excel and Word, letting AI agents handle tasks via natural language.
- OpenAI added safety routing and parental controls to ChatGPT, balancing protection with concerns over user freedom.
- AI is transforming ransomware into a systemic, precision-driven threat, forcing organizations to adopt advanced defenses beyond traditional security.
- AI actress Tilly Norwood went viral, attracting agencies but sparking backlash over replacing human actors.
TRENDING
Daily AI Launches
- Integrity launched a “unified project brain” linking notes and canvases with AI chat for context-grounded co-creation.
- The Nexa SDK has launched as a universal toolkit for running any AI model locally on any device, with support for CPU, GPU, and specialized NPUs.
- Deamoy has launched a fast AI app builder that creates editable sites from a single prompt and keeps visual edits and AI prompts in perfect sync.
- Fern has launched a tool that automatically generates multi-language SDKs and branded, interactive documentation from a single API specification.
Featured AI Tools
- 🤩 Pine AI Agent recovers what’s rightfully yours.
- 💬 CREAO makes future apps possible for all.
- 📚 Heardly is the Fast Way to read Best Book.
- 🪶 CopyOwl is the First AI Research Agent, deep research on any topic in one click.
TOGETHER WITH US
AI Secret Media Group is the world’s #1 AI & Tech Newsletter Group, boasting over 1 million readers from leading companies such as OpenAI, Google, Meta, and Microsoft. Our Newsletter Brands:
- AI: AI Secret
- Tech: Bay Area Letters
We've helped promote over 500 Tech Brands. Will yours be the next?
Email our co-founder Mark directly at [email protected] if the button fails.
Latest Daily Rundowns
MagnaMagnaMagnaMore AI Stories
Ben
Magna
Magna
Good Morning, AI Enthusiasts!
New versions drop, new rules emerge, and what looked solid a quarter ago now feels like scaffolding for something else entirely.
LLM
The Turing Laureate Who Thinks Big AI Got Dumb
📌 What’s happening: Richard Sutton—reinforcement learning pioneer, “Bitter Lesson” prophet, and Turing Award winner—says today’s AI boom is off track. Instead of piling tokens into trillion-param LLMs, Sutton pushes Oak (Options and Knowledge): agents that learn by doing, not by memorizing the internet. His recipe: continual learning, world models, planning, and reward-driven abstraction.
🧠 How this hits reality: LLMs are front-loaded with static knowledge; Sutton’s Oak starts empty, builds from scratch, and never stops learning. That means less catastrophic forgetting, more generalizable skills. For labs betting billions on GPU-fuelled pretraining, his critique is existential: they’re optimizing for demos, not durable intelligence.
🛎️ Key takeaway: Sutton’s not just throwing shade—he’s saying the LLM empire is architected on sand, while real intelligence demands roots that grow, not weights that bloat.
TOGETHER WITH AMPLITUDE
Meet the AI-Powered Digital Analytics Leader
Unlock the next era of product intelligence with Amplitude—a powerful analytics and optimization platform built to understand user behavior, anticipate needs, and drive intelligent action.
Amplitude transforms user data into actionable insights—from clicks and scrolls to session flows and feature interactions, Amplitude reveals how users engage and where they drop off.
BANKING
Casca AI Rewires Small-Bank Lending
📌 What’s happening: Casca just pulled in $29M Series A to scale its AI-native loan origination platform. The pitch: process loans 10x faster than fintechs, 30x faster than banks, while cutting 20 hours of grunt work per officer. Local banks suddenly get fintech speed without fintech burn.
🧠 How this hits reality: This isn’t “AI for efficiency”—it’s a moat-breaker. By automating credit checks, document intake, and even customer conversations, Casca hands community banks the weapons they need to keep borrowers from defecting to JPMorgan or Stripe Capital. The back-office bottleneck—where most banks quietly bleed customers—just got industrially crushed.
🛎️ Key takeaway: Casca isn’t selling software—it’s arming small banks with fintech-grade artillery. And once borrowers taste instant credit, nobody’s waiting three weeks again.
NEW LAUNCH
DeepSeek Drops V3.1
📌 What’s happening: DeepSeek just pushed V3.1 live: bigger context, same mystery box. No benchmarks, no routing system, no multimodal flex — just “look, it reads more tokens now.” Meanwhile GPT-5 and Claude 4.1 are shipping orchestration layers, hybrid reasoning, and enterprise-grade 200K–1M windows.
🧠 How this hits reality: For founders and operators, this isn’t a playbook rewrite — it’s a lagging feature check. Teams betting on DeepSeek won’t get the agentic tooling, safety rails, or enterprise integrations rivals already normalized. Translation: if you’re deploying at scale, you’re still buying OpenAI or Anthropic licenses while DeepSeek plays catch-up.
🛎️ Key takeaway: DeepSeek isn’t redefining the curve — it’s studying last semester’s syllabus.
Anthropic: DeepSeek Drops V3.1
No benchmarks, no routing system, no multimodal flex — just “look, it reads more tokens now.” Meanwhile GPT-5 and Claude 4.1 are shipping orchestration layers, hybrid reasoning, and enterprise-grade 200K–1M windows.
DeepSeek Drops V3.1
Teams betting on DeepSeek won’t get the agentic tooling, safety rails, or enterprise integrations rivals already normalized. Translation: if you’re deploying at scale, you’re still buying OpenAI or Anthropic licenses while DeepSeek plays catch-up.
QUICK HITS
- Meta launches an AI-powered translation and lip-syncing feature for creators on Instagram and Facebook, starting with English and Spanish.
- DeepSeek-R1 is losing momentum in India as it struggles with a lack of local language support, operational complexity, and intense competition.
- A majority of Californians are concerned about the future of AI and believe the technology is progressing too quickly, a new poll finds.
- A new study suggests doctors' ability to spot colon abnormalities declines after using AI assistance, raising concerns about over-reliance on the technology.
- The AI talent race is splitting, with big tech buying top experts while startups prioritize recruiting for cultural fit and adaptability.
TRENDING
Daily AI Launches
- April launched a voice AI executive assistant for hands-free management of your email and calendar.
- ElevenLabs released its Eleven Music API for integrating high-quality AI music into any product or workflow.
- Chance AI launched an Android app that acts as a visual agent, using your camera to provide meaningful insights about the world.
- AutonomyAI launched Fei, an autonomous AI engineer that writes production-grade code following your team's existing codebase and standards.
Featured AI Tools
- 📚 Heardly is the Fast Way to read Best Book.
- 🪶 CopyOwl is the First AI Research Agent, deep research on any topic in one click.
- 🦾 Flot AI writes, reads, and remembers across any apps and webs.
TOGETHER WITH US
AI Secret Media Group is the world’s #1 AI & Tech Newsletter Group, boasting over 1 million readers from leading companies such as OpenAI, Google, Meta, and Microsoft. Our Newsletter Brands:
- AI: AI Secret
- Tech: Bay Area Letters
- Futuristic: Posthuman
We've helped promote over 500 Tech Brands. Will yours be the next?
Email our co-founder Mark directly at [email protected] if the button fails.
Latest Daily Rundowns
MagnaMagnaMagnaMore AI Stories
Magna
MagnaBen
Good Morning, AI Enthusiasts!
In a move that can only be described as the AI equivalent of ghosting your date after a third dinner, Anthropic left Windsurf scrambling like a cat on a hot tin roof.
SPONSORED BY
QUICK HITS
- Microsoft has launched a free AI video generator in Bing to challenge Google's dominance in content creation.
- Google DeepMind is developing an AI-powered "next-generation email" system to automate email management and reduce inbox clutter.
- US labor unions are battling against job displacement from AI advancements, facing political and corporate challenges in securing worker protections.
- Mistral has launched Codestral Embed, a new code embedding model that surpasses OpenAI and Cohere in real-world retrieval tasks.
- Amazon is investing $10 billion in new data centers in North Carolina to enhance its AI infrastructure.
WAR
The Tale of Windsurf, Anthropic, and the Battle of the AI Titans
In a plot twist worthy of a Silicon Valley drama series, Windsurf, the AI-powered coding startup recently acquired by OpenAI for a jaw-dropping $3 billion, found itself in the crosshairs of Anthropic, a rival AI powerhouse.
- Windsurf, having achieved impressive growth with its innovative coding platform, faced an unexpected setback when Anthropic cut off its access to key AI models shortly after being acquired by OpenAI for $3 billion.
- With a mere five days to adapt, Windsurf quickly sought to implement third-party inference services and introduced a workaround for users to maintain some model access, despite criticism that it was a costly and complicated solution.
- Anthropic’s decision has triggered discussions about the implications of such competitive strategies, with concerns about potential harm to innovation and user reliance on these technologies.
What's interesting is that Windsurf opts for pricier third-party solutions to access Anthropic's advanced model capabilities instead of using OpenAI's models. What better way to say “I trust you” than to pay for it?
OpenAI: The Tale of Windsurf, Anthropic, and the Battle of the AI Titans
In a plot twist worthy of a Silicon Valley drama series, Windsurf, the AI-powered coding startup recently acquired by OpenAI for a jaw-dropping $3 billion, found itself in the crosshairs of Anthropic, a rival AI powerhouse.
The Tale of Windsurf, Anthropic, and the Battle of the AI Titans
In a plot twist worthy of a Silicon Valley drama series, Windsurf, the AI-powered coding startup recently acquired by OpenAI for a jaw-dropping $3 billion, found itself in the crosshairs of Anthropic, a rival AI powerhouse.
TOGETHER WITH MACROCOSMOS
Pretraining is the hardest part of AI to decentralize - until now
Vast data centres, enormous compute, and insatiable energy demands: only big tech can train LLMs. Even open-source models like Llama or Qwen are centrally trained, leaving DeAI dependent on centralized competitors.
Macrocosmos, an open-source AI lab on Bittensor, just launched IOTA - Incentivized Orchestrated Training Architecture. It allows anyone, anywhere to participate in LLM training-at-home:
✅ No memory constraints = wider accessibility.
✅ Fair, real-time rewards = co-ownership with proportionate attribution.
✅ 128× compression + trustless weight merging = global scale, fast.
Check out our interactive 3D dashboard. Join now to contribute compute, get rewarded, and train LLMs - and help make data centres a thing of the past.
COPYRIGHT
DeepSeek Might Have Used Google's Gemini To Train Its Newest Model
Chinese AI lab DeepSeek is alleged use of Google's Gemini data to train its latest AI model, R1-0528. The scrutiny over the data sources used for training AI models raises ethical concerns about data ethics, competition, and the integrity of AI outputs.
- DeepSeek has been accused of using outputs from Google's Gemini to train its R1-0528 model, with researchers suggesting similarities in language and expression between the two.
- This is not the first time DeepSeek has faced scrutiny; previous models were found to identify themselves as ChatGPT, hinting at potential training on OpenAI’s datasets.
- OpenAI has noted that DeepSeek may have employed distillation techniques, which involve extracting data from more advanced AI models, to enhance its own offerings.
The implications of such data sourcing accusations are profound for the AI landscape. In response to such practices, AI companies like OpenAI and Google are tightening security measures to prevent unauthorized training on their models.
OpenAI: DeepSeek Might Have Used Google's Gemini To Train Its Newest Model
This is not the first time DeepSeek has faced scrutiny; previous models were found to identify themselves as ChatGPT, hinting at potential training on OpenAI’s datasets.
DeepSeek Might Have Used Google's Gemini To Train Its Newest Model
Chinese AI lab DeepSeek is alleged use of Google's Gemini data to train its latest AI model, R1-0528. The scrutiny over the data sources used for training AI models raises ethical concerns about data ethics, competition, and the integrity of AI outputs.
DeepSeek Might Have Used Google's Gemini To Train Its Newest Model
Chinese AI lab DeepSeek is alleged use of Google's Gemini data to train its latest AI model, R1-0528. The scrutiny over the data sources used for training AI models raises ethical concerns about data ethics, competition, and the integrity of AI outputs.
TRENDING TOOLS
- 🗒️ Connected Papers explores connected papers in a visual graph.
- 🎞️ HeyEditor offers multiple AI video and photo editing tools.
- 📛 Cleanvoice AI edits your podcast in 10 mins automatically.
- ☎️ Chaterimo is an AI-powered customer service chatbot.
- 🎥 Rewin AI gets viral Youtube and TikTok scripts in seconds.
- 🪶 CopyOwl is the First AI Research Agent, deep research on any topic in one click. (Featured)
- 📚 Heardly is the Fast Way to read Best Book. (Featured)
- 🦾 Flot AI writes, reads, and remembers across any apps and webs. (Featured)
1 Million+ AI enthusiasts are eager to learn about your product.
AI Secret is the world’s #1 AI Newsletter, boasting over 1 million readers from leading companies such as OpenAI, Google, Meta, and Microsoft. We've assisted in promoting Over 500 AI-Related Products. Will yours be the next?
What We Can Offer:
- Launch an Advertising Campaign
- Introduce New Product or Features
- Other Business Cooperation
Email our co-founder Mark directly at [email protected] if the button fails.
Magna
Magna
Magna
Good Morning, AI Enthusiasts!
In the age of intelligent machines, the art of minimalism is redefined.
SPONSORED BY
QUICK HITS
- Opera has launched Opera Neon, the first AI agentic browser designed to understand user intent and automate web tasks.
- Tesla plans to launch its robotaxi service in Austin on June 12, 2025.
- Mistral has launched its new code embedding model, Codestral Embed, which outperforms competitors like OpenAI and Cohere in real-world retrieval tasks.
- Nvidia reported a 69% revenue increase year-over-year, reaching $44.1 billion for Q1 2025, despite facing export licensing challenges.
- The U.S. Department of Justice has indicted 16 defendants linked to DanaBot, a Russian malware platform responsible for significant cybercrime, following its recent takedown facilitated by agentic AI advancements.
TECH
Meta Study Shows Shorter Reasoning Boosts AI Accuracy by 34%
The recent study by Meta's FAIR team and The Hebrew University of Jerusalem has unveiled a revolutionary concept in AI development: shorter reasoning chains can enhance the accuracy and efficiency of AI systems. This paradigm shift challenges the conventional wisdom of lengthy, detailed reasoning processes, suggesting that sometimes, less truly is more.
- The AI industry traditionally believed that longer, step-by-step reasoning leads to better performance, a notion supported by major AI labs. However, this approach is costly, time-consuming, and prone to errors.
- Researchers found that shorter reasoning chains resulted in a 34.5% increase in accuracy while reducing computational costs by 40%. The innovative "short-m@k" method executes parallel reasoning attempts, stopping computations early and selecting the best answer through majority voting.
- The short-m@k approach promises significant cost savings for companies, improving accessibility to AI development and making real-time applications like chatbots and virtual assistants faster and more efficient while also reducing the environmental impact of AI training.
This groundbreaking study heralds a new era for AI development, emphasizing that efficiency can triumph over complexity. By adopting shorter reasoning chains, companies can not only save billions on computational resources but also create faster, more user-friendly AI applications.
Meta Study Shows Shorter Reasoning Boosts AI Accuracy by 34%
The recent study by Meta's FAIR team and The Hebrew University of Jerusalem has unveiled a revolutionary concept in AI development: shorter reasoning chains can enhance the accuracy and efficiency of AI systems. This paradigm shift challenges the conventional wisdom of lengthy, detailed reasoning processes, suggesting that sometimes, less truly is more.
TOGETHER WITH CLICKUP
The Everything App for Work
If you've ever dreamed of a magical place where your tasks, projects, and to-do lists come together like a perfectly orchestrated symphony, then you’ve hit the jackpot with ClickUp.
ClickUp is the ultimate Swiss Army knife of productivity, designed to slice through chaos and serve up success on a silver platter. Whether you’re juggling spreadsheets, chasing deadlines, or trying to keep your team on the same page without resorting to smoke signals, ClickUp has got you covered.
NEW LAUNCH
DeepSeek R1 Model Update: A "Minor" Upgrade with Major Implications
DeepSeek, has recently launched a significant upgrade to its reasoning AI model, the DeepSeek-R1. Despite being labeled a "minor" update, this new version showcases impressive advancements that position it as a formidable competitor to industry giants such as Google.
- The R1-0528 upgrade boasts deep reasoning abilities comparable to Google’s Gemini 2.5 Pro. This remarkable improvement in mathematical reasoning and reduced hallucination rates positions R1-0528 as a serious contender in the reasoning AI sphere.
- Evaluators have praised the R1-0528 for its upgraded writing skills, producing text that is more natural and emotionally resonant. Its literary quality now rivals that of leading models, demonstrating a significant leap in open-source AI writing potential.
- The model’s fast and thoughtful reasoning style allows it to tackle complex prompts with enhanced accuracy, making it a versatile tool for developers and businesses across various applications.
- The ability of R1-0528 to engage in long-term thinking for tasks lasting 30-60 minutes opens new avenues for strategic planning and complex problem-solving, enhancing its utility in diverse fields.
The launch of the R1-0528 under an MIT license is a revolutionary step that challenges the established dominance of closed-source AI models. The enhanced capabilities of R1-0528 provide a cost-effective solution for businesses, allowing them to leverage AI without exorbitant expenses. By making advanced capabilities accessible, DeepSeek is democratizing AI technology, fostering innovation, and inspiring other players in the open-source community.
Google: DeepSeek R1 Model Update: A "Minor" Upgrade with Major Implications
The R1-0528 upgrade boasts deep reasoning abilities comparable to Google’s Gemini 2.5 Pro. This remarkable improvement in mathematical reasoning and reduced hallucination rates positions R1-0528 as a serious contender in the reasoning AI sphere.
DeepSeek R1 Model Update: A "Minor" Upgrade with Major Implications
DeepSeek, has recently launched a significant upgrade to its reasoning AI model, the DeepSeek-R1. Despite being labeled a "minor" update, this new version showcases impressive advancements that position it as a formidable competitor to industry giants such as Google.
TRENDING TOOLS
- 🧱 Protectstar provides secure AI meeting military-grade standards.
- 🎯 AdCreative AI is an powerful AI tool for advertising.
- 📲 SocialPostGPT creates engaging social media content with AI.
- 🧾 TaxTim is your ultimate tax companion.
- 🏪 Storebuild.ai creates your online store in just 5 minutes.
- 🪶 CopyOwl is the First AI Research Agent, deep research on any topic in one click. (Featured)
- 📚 Heardly is the Fast Way to read Best Book. (Featured)
- 🦾 Flot AI writes, reads, and remembers across any apps and webs. (Featured)
1 Million+ AI enthusiasts are eager to learn about your product.
AI Secret is the world’s #1 AI Newsletter, boasting over 1 million readers from leading companies such as OpenAI, Google, Meta, and Microsoft. We've assisted in promoting Over 500 AI-Related Products. Will yours be the next?
What We Can Offer:
- Launch an Advertising Campaign
- Introduce New Product or Features
- Other Business Cooperation
Email our co-founder Mark directly at [email protected] if the button fails.
Magna
Magna
Magna
🛎️ Mind Captioning
Plus: Whistleblowers to Job Loss, OpenAI Launched GPT-5.1
🛎️ DeepSeek Drops OCR Bomb
Plus: Meta Bans Competitors, Claude for Life Sciences
🛎️ DeepSeek Freefall
Plus: 7-11 Becomes Robotics Lab, Chief of Staff
🛎️ Sand Empire
More Stories: AI Rewires Lending, DeepSeek Drops V3.1
🛎️You Are Banned
DeepSeek accused of using Google's Gemini to train its latest model.
DeepSeek Might Have Used Google's Gemini To Train Its Newest Model
In the ever-evolving world of artificial intelligence, where breakthroughs are as common as coffee breaks, Chinese AI lab DeepSeek has dropped a bombshell—or rather, a
🛎️Less Is More
DeepSeek R1 Model Update: A "Minor" Upgrade with Major Implications
DeepSeek R1 Model Update: A "Minor" Upgrade with Major Implications
DeepSeek, the ambitious Chinese AI startup, has recently rolled out a "minor" upgrade to its R1 reasoning AI model, officially dubbed the DeepSeek-R1-0528. But
A Big Issue: As AI Improves, Its Hallucinations Increase
Artificial Intelligence (AI) is at a crossroads. The past two years have seen an explosion in the capabilities of generative AI models, especially large language models
