[AI DAILY NEWS RUNDOWN] The Mythos Cyber-Leak, the MIT Layoff Autopsy, and Meta’s Open-Source Brain (March 27th 2026)

Meta just open-sourced TRIBE v2, an AI model trained on brain scans from 700+ people that simulates neural activity across vision, hearing, and language — with its synthetic predictions actually outperforming real fMRI recordings.

The details:

• Trained on 1,000+ hours of brain data, v2 leaps from 1,000 brain regions to 70,000, with 700+ subjects up from just 4 volunteers in the original.

• TRIBE v2’s predictions matched population-level brain activity better than most real scans, which often get clouded by heartbeats, movement, and noise.

• The team replicated decades of neuroscience findings in software, correctly pinpointing brain regions for faces, speech, and text with zero scans.

• Meta open-sourced the code, weights, and a live demo, letting any researcher start running virtual brain experiments without building from scratch.

Why it matters: Neuroscience has long required putting people inside expensive scanners for every new experiment, a bottleneck that’s kept entire fields moving one study at a time. TRIBE v2 could do for brain research what AlphaFold did for protein structure: compress months of scanning into seconds of compute.

Apple to unlock Siri for rival AI assistants

Image source: Apple

The Rundown: Apple plans to open up the upcoming Siri revamp for other models starting with iOS 27, according to Bloomberg — ending ChatGPT’s exclusive integration and letting users choose which AI handles their queries directly from the assistant.

The details:

• Users will be able to pick their preferred AI in ‘extensions’ settings and route questions to models of their choice via Siri with the incoming iOS 27.

  
  
  

• ChatGPT is currently the only model compatible with Siri commands via its 2024 deal, but use of that integration has reportedly been ‘minimal’.

• Bloomberg said chatbots in the App Store could also be a revenue stream, with Apple taking a cut of AI subscriptions purchased across its devices.

• Apple is expected to introduce the new Siri AI overhaul powered by Gemini at its WWDC developer event in early June.

Why it matters: Google is already rebuilding Siri’s underlying tech with Gemini, and ChatGPT has had a spot since 2024. Now, Apple is letting the rest of the field in to provide more user choice. It’s a smart move — skip the model war entirely, layer the best AI on top of a billion iPhones, and let its hardware moat do the rest.

OpenAI pauses erotic chatbot plans indefinitely

• OpenAI has paused its plans to launch an erotic “adult mode” for ChatGPT indefinitely, confirming to the Financial Times that it is shifting focus toward its core products instead.

• The company wants more time to research the potentially harmful effects of sexually explicit chats and the emotional attachments they may create, while investors also weren’t excited about the project.

• This is the second major product OpenAI pulled back this week, after discontinuing its Sora AI video-generation app to redirect compute power to other higher-priority projects going forward.

Apple may build smaller AI models from Gemini

• Apple has gained full access to Google’s Gemini model and plans to distill it into smaller models that can run directly on Apple devices without an internet connection.

• The distillation process works by feeding Gemini’s high-quality answers and reasoning information into smaller, cheaper models that learn its internal computations while requiring less computing power.

• Apple is building a smarter, chatbot version of Siri for iOS 27 using Gemini, but has hit issues because Gemini was tuned for chatbot and coding tasks that don’t always match Apple’s needs.

Quantum computers could break encryption by 2029, warns Google

• Google published a formal plan to move all of its infrastructure to post-quantum cryptography by 2029, warning that quantum computers capable of breaking current encryption may arrive sooner than expected.

• The company highlighted “harvest now, decrypt later” attacks as an already active threat, where bad actors steal encrypted data today planning to unlock it once quantum machines become powerful enough.

• Over 6.8 million Bitcoin sitting in vulnerable addresses could eventually be at risk, but Bitcoin developers have started work on quantum-resistant upgrades through BIP 360, a new address format proposal.

Google TurboQuant cuts AI memory use by 6x

• Google Research announced TurboQuant, a new compression algorithm that can reduce AI working memory — known as the KV cache — by at least 6x without losing performance or accuracy.

• The method combines two techniques called PolarQuant and QJL, using vector quantization to clear cache bottlenecks, and the team plans to present their findings at ICLR 2026 next month.

• TurboQuant is still a lab breakthrough and only targets inference memory, not training, so it wouldn’t solve the wider RAM shortages even if successfully deployed in real-world systems.

Reddit to require human verification for suspicious accounts

• Reddit will now force accounts flagged for suspicious behavior to verify they are human, using passkeys, biometrics, and bot labeling as the platform removes around 100,000 automated accounts every day.

• CEO Steve Huffman said passkeys serve as a baseline check but cannot prove individuality, while biometric options like World ID’s iris-scanning and Face ID offer stronger proof but raise privacy concerns.

• Co-founder Alexis Ohanian expressed skepticism about selling face-scanning to Redditors, highlighting tension between the platform’s pseudonymous culture and the technical demands of proving personhood at scale.

Wikipedia bans AI from writing its articles

Image source: Wikipedia

Wikipedia’s volunteer editors banned the use of AI to write articles on the foundation’s English-language site, a move the policy’s author called a “pushback against enshittification and forceful push of AI by so many companies”.

The details:

• Prior attempts at broad AI rules failed to reach consensus, but mounting AI-generated errors pushed editors to a near-unanimous 40-2 vote.

• The ban covers writing or rewriting articles with LLMs, with editors still allowed to use AI for grammar fixes and translations with human review.

• The policy’s author said the change could “spark a broader change” and “empower communities on other platforms” to set AI rules on their own terms.

• StackOverflow and German Wikipedia have enacted similar bans, with Spanish Wikipedia going further to fully ban the use of AI, even for editing purposes.

Why it matters: AI text reportedly surpassed human output for the first time in 2025, and Wikipedia is trying to hold the human line, all while Elon pushes Grokipedia (an AI-created version of Wikipedia) in the exact opposite direction. The internet’s most-used knowledge base bet against the current, but how long that holds is anyone’s guess.

Open-Source Speed Demon

Nvidia, the dominant supplier of AI chips, released a competitive open-source large language model whose speed tops its size class — the first open-weights leader to come from the United States since last year, when Meta delivered Llama 4.

What’s new: Nvidia released Nemotron 3 Super 120B-A12B, a large language model designed for agentic applications, including not only weights but also training datasets and recipes. It is the second in a planned family of three: Nvidia released Nemotron 3 Nano-39B-A3B in December 2025, and Nemotron 3 Ultra-500B-A50B is forthcoming.

• Input/output: Text in (up to 1 million tokens), text out (up to 1 million tokens)

• Knowledge cutoff: June 2025 (pretraining data), February 2026 (fine-tuning data)

• Architecture: Hybrid mamba-2/transformer/mixture-of-experts with multi-token prediction layers (120 billion parameters, 12 billion active per token)

• Training data: 25 trillion tokens of curated data scraped from the web and synthesized in 20 natural languages and 43 programming languages

• Features: Tool calling, structured outputs, seven languages (Chinese, English, French, German, Italian, Japanese, Spanish), reasoning modes (off, low, regular)

• Performance: Fastest open-weights model of its size (442 output tokens per second), leads open-weights models on PinchBench test of agentic tasks

• Availability/price: Weights and datasets free to download under a license that permits noncommercial and commercial uses (rights terminate if safety guardrails are removed without replacement or if the user files patent or copyright litigation against Nvidia), free chat via Nvidia and OpenRouter, API around $0.30/$0.80 per 1 million tokens of input/output via third-party providers

How it works: Nemotron 3 Super’s hybrid architecture interleaves mamba-2, attention, and modified MoE layers with multi-token prediction heads that generate a number of tokens per forward pass.

Behind the news: Nvidia plans to invest $26 billion over five years to develop open-weights models — a substantial commitment. The announcement coincides with shifts in the open-weights landscape that could affect Nvidia’s business. Chinese companies, including Alibaba, Moonshot AI, and Z.ai, lately have built the most capable open-weights models, and they are building alternatives to Nvidia GPUs and Cuda software. For instance, DeepSeek has reportedly trained an upcoming model entirely on Huawei’s Ascend chips and Cann software.

Why it matters: Nemotron 3 Super gives developers a fast, fully open model for agentic applications, with training data, recipes, and tools alongside the weights. This openness also serves Nvidia’s business goals. Chinese open-weights models are growing more capable and increasingly streamlined to run on non-Nvidia chips, creating a risk that developers who previously relied on Nvidia will look elsewhere. Nemotron gives them a reason not to.

We’re thinking: Who better to optimize a model for GPUs than the company that designs the GPUs? From custom numerical formats to inference software, Nvidia can co-design hardware and software in ways that few model developers can match. Nvidia is betting that building models will help sell chips and vice versa.

The “AI is replacing software engineers” narrative was a lie. MIT just published the math proving why. And the companies who believed it are now begging their old engineers to come back.

Since 2022, the tech industry has been running a coordinated narrative.

AI will replace 80 to 90% of software engineers. Learning to code is pointless. Developers are obsolete. but what if i tell you that It wasn’t a prediction. It was a headline designed to create fear. And it worked on millions of students and engineers who genuinely believed their careers were over before they started.

It’s 2026 now. Let’s look at what actually happened.

In 2025, 1.17 million tech workers were laid off. Everyone said it was AI. Companies said it was AI. The news said it was AI.

You want to know what percentage of those people actually lost their jobs because AI automated their work?…5%, I’m not lying atp, its literally around 5%, 55k people out of 1.17 million. That’s it.

And according to an MIT study, nearly 95% of companies that adopted AI haven’t seen meaningful productivity gains despite investing millions. The revolution that was supposed to make engineers obsolete couldn’t even pay for itself.

now coming to the main point, So if AI didn’t cause the layoffs, what did?

Source : https://arxiv.org/pdf/2505.10465

What Else Happened in AI on March 27th 2026?

Google rolled out Gemini 3.1 Flash Live, a new voice AI with upgrades in speed, task completion, and realism, to power convos across Search, Gemini Live, and its API.

Mistral released Voxtral TTS, a lightweight voice AI that clones any speaker from a 3-second clip and generates natural-sounding speech across 9 languages.

OpenAI has reportedly shelved its planned erotic chatbot mode indefinitely after pushback from staff and investors.

Novo Nordisk is deploying AI agents across clinical trial ops, with the pharma giant saying the tech is trimming approval timelines and reducing the need for contractors.

Suno launched v5.5 of its AI music generator, adding voice cloning, custom model tuning, and personalized style learning for Pro subscribers.

Cohere released Transcribe, a free open-source speech recognition model that tops HuggingFace’s accuracy leaderboard across 14 languages — taking the No. 1 spot.

Claude AI Maker Anthropic Considers IPO as Soon as October.

Meta Releases TRIBE v2: A Brain Encoding Model That Predicts fMRI Responses Across Video, Audio, and Text Stimuli.

Tencent AI Open Sources Covo-Audio: A 7B Speech Language Model and Inference Pipeline for Real-Time Audio Conversations and Reasoning.

LISTEN DAILY ADS-FREE at: