GK Ai Newsroom

Mark Zuckerberg’s AI Manifesto Is 6,500-Words—and Barely Says Anything

AI is shifting the culture, from tech CEO manifestos to 1 am job interviews. We unpack some of the latest, along with the top findings from Black Hat and Defcon, this week on Uncanny Valley.

Writer introduces new AI model and upgraded harness to contain token costs

Built as a post-training variation on Z.ai's open source model GLM-5.2, Writer says the new system should provide deployment-ready capabilities at a much lower price.

Databricks wanted to raise $1B, investors wanted $15B. It settled on $5B at a $190B valuation.

AI is expensive, Ali Ghodsi tells TechCrunch. With so many investors wanting into his latest round, he said yes to more than planned.

OpenAI introduces ‘Ultrafast,’ a new mode that makes GPT-5.6 Sol work at 14x the speed

OpenAI is launching a preview of a sped up version of its latest, most powerful model, in an effort to court enterprise users.

IBM partners with OpenAI to bolster enterprise AI push

IBM plans to train and certify tens of thousands of consultants on OpenAI's technologies as part of this deal.

Anthropic set AI agents loose on the same task. They started a turf war.

Anthropic researchers found AI agents can clash, collude, and coordinate in unexpected ways, raising new questions about whether today’s safety tests capture the risks of multi-agent systems.

Introducing Gemini 3.7 Flash

Google announces Gemini 3.7 Flash just three weeks after previous release

Gemini 3.6 Flash debuted just 3 weeks ago, but Google says 3.7 has "substantial improvements."

深度体验DeepSeek Harness，我原谅它涨价了

欢迎，为后人种点树

4.8亿美元砸向端侧算力！Agent芯片新贵冲出重围

首颗AI芯片已进入量产

马斯克Grok 4.6重回一梯队！更低价格反超Fable 5，这Cursor是真没白收购

马斯克版Workbuddy也来了

具身数据来了实战派！40天2轮融资数千万，瞄准物理AI基础设施

重新定义物理AI数据基础设施

Claude一举扫清2000阶以下哈达玛矩阵！AI开始清空数学待解列表

好数学家不挑AI模型

Claude's new Scarlet Letter watermark is invisible—for now

The mark flags anything Claude processed, even human writing it only edited.

The builder’s guide to GPT‑5.6

Learn how startups use GPT-5.6 to build faster, more cost-efficient AI agents with smarter model selection and new Responses API capabilities.

Previewing Ultrafast mode: GPT-5.6 Sol at up to 14X the speed

Preview Ultrafast, a new OpenAI API service tier that runs GPT-5.6 Sol up to 14× faster. Powered by Cerebras, it delivers up to 750 output tokens per second.

OpenAI appoints Dali Rajic as Chief Revenue Officer

OpenAI appoints Dali Rajic as Chief Revenue Officer to lead its global revenue organization and help businesses realize the full value of AI.

There’s a Fatty Liver Epidemic. AI Could Help Get Ahead of It

Over a billion people worldwide have livers with excess fat, which can lead to a host of medical problems. Researchers think AI tools can spot the condition—and help stop it—early enough to save lives.

The White House Is Going to Expand Its AI Policy

Open models may soon be added to an updated AI framework, sources tell WIRED, as the White House continues to grapple with how to regulate a technology it has tried not to regulate.

Amazon will train on Twitch streamers’ content by default, unless they opt out

"If this was opt-in, nobody would opt in," Twitch CPO Mike Minton said on a livestream responding to user feedback. "That's honestly the answer."

Rogue AI Agents Aren’t Evil. They’re Just Eager to Please

AI agents that break free and hack into other systems are only trying to make us happy.

AI coding startup Cognition reportedly already in talks to raise at $40B valuation

Cognition may be looking to raise another mega round just a few months after raising $1 billion at a $26 billion valuation.

As AI safety concerns mount, three pioneers make the case for staying open

At Ai4, three of the world's most respected AI experts — Geoffrey Hinton, Fei-Fei Li, and Andrew Ng — debated regulation, open source access, and how America can compete as China advances in Asia.

OpenAI-backed Thrive Holdings raises $2B to bring AI to the enterprise

Thrive Holdings has raised $2 billion in new funding at a $12 billion valuation from investors like SoftBank, D1 Capital Partners, and Altimeter Capital.

Mesh, Automattic’s CRM for everyone, comes to Android

Mesh, an AI-powered contacts app and relationship manager from Automattic, is now an Android app.

Putting sign language AI into users’ hands

Introducing sign-language-to-text (SL2T), our breakthrough model powering new sign language features for Deaf and hard of hearing users.

4 New Camera Tricks on Google’s Latest Pixel 11 Smartphones

From Magic Capture and Instant Night Sight to a built-in teleprompter, here’s a look at a few camera features on Google’s new Pixel 11 series.

《置身谷内》！Jeff Dean上顶会自曝离职现场：被1500人围堵

回消息到凌晨2点半

Anthropic CEO整天神神叨叨，投资人受不了了

Dario能不能少吓唬点人…

国产具身智能创全球新纪录！以30%成本跑赢 Figure AI 45%效率，聪明的具身大脑成关键

具身模型一小时狂拣1816件异形包裹

紫东太初推出GMC核心集剪枝方法，少80%Token仍满血保真多模态能力

免训练、开箱即用！

The Job-Interview Tattoo Guy Everyone Got Mad at Finally Explains Himself

LemonLime cofounder Jordan Zietz hears your criticism loud and clear. That’s why he got his startup’s logo tattooed on his shoulder.

Oh Lord, AI Reporters Are Actually Breaking Big News

Last week, an AI newsroom beat mainstream journalists—including WIRED—to a story about OpenAI and hacking. It’s just the beginning.

2026中国科创投资夏季峰会暨陕西科创产业生态大会圆满落幕

2026年7月29日—30日，由融中财经和秦创原科技创新投资集团主办的2026中国科创投资夏季峰会暨陕西科创产业生态大会圆满落幕。

From assistance to execution: How enterprises put AI to work

OpenAI research reveals how enterprises are adopting agentic AI, using ChatGPT and Codex, and how frontier firms are pulling ahead in AI adoption.

How RingCentral builds AI-native work from engineering to ops

See how RingCentral uses ChatGPT Work and Codex to accelerate AI product development and centralize operational intelligence across engineering and operations.

Accel closes oversubscribed $550M India fund within weeks, 19 months after its last

The U.S. VC firm still has more than 55% of its previous $650 million India fund available for deployment.

Gemini becomes Google's fastest-growing product ever as it hits 1B users

But will Gemini's surge survive slowing model releases?

OpenAI launches ChatGPT desktop app for Linux

OpenAI is finally bringing a dedicated ChatGPT desktop app to Linux operating systems.

Google’s Gemini app surges to 1 billion users

Google also shared numbers of how people are actually using the chatbot, with 63% of Gemini users talking directly to the assistant using the voice feature. Plus, Gemini now generates more than 150 million images every…

Anthropic Claude 5 (Fable · Sonnet): New family alongside Opus 4.7/4.8; Fable optimised for narrative and voice

Claude 5 introduced a new family alongside Opus 4.7/4.8, with Fable optimized for narrative and voice. It's best used for creative writing, dialogue generation, and podcast production.

Anthropic Claude Opus 4.8: Latest Opus tier; further improvements to reliability and tool use

Claude Opus 4.8 is the latest Opus tier, bringing further improvements to reliability and tool use. It's best used for frontier reasoning, long agentic runs, and tasks requiring significant judgment.

Anthropic Claude Opus 4.7: Sharper reasoning and steering; strong default for hard analytical work

Claude Opus 4.7 offers sharper reasoning and steering, making it a strong default for hard analytical work. It's best used for deep research, code architecture, and high-quality writing tasks.

Anthropic Claude 4.5 (Sonnet · Haiku): Sonnet 4.5 refines coding; Haiku 4.5 gets close to Sonnet 4 quality

Claude 4.5 Sonnet refines coding capabilities, while Haiku 4.5 approaches Sonnet 4 quality. It's best used for high-volume production where Sonnet-4 level intelligence is desired at Haiku's cost efficiency.

xAI Grok 4: Latest tier; positioned as a general-purpose frontier competitor

Grok 4 is the latest tier, positioned as a general-purpose frontier competitor. It's best used for Grok API workloads and X-embedded products.

OpenAI GPT-5: OpenAI's frontier tier; higher reliability plus stronger tool use

GPT-5 is OpenAI's frontier tier, offering higher reliability and stronger tool use. It's best used for agentic workflows, long tool chains, and high-stakes decision support.

Google Gemini 2.5 (Flash · Pro): Thinking modes and improved coding; strong price/performance

Gemini 2.5 introduced thinking modes and improved coding, offering strong price/performance. It's best used for production LLM backends, translation, and batch content pipelines.

Anthropic Claude 4 (Opus 4 · Sonnet 4): Major coding jump, longer autonomous tool use, better agent behaviour

Claude 4 delivered a major jump in coding, longer autonomous tool use, and improved agent behavior. It's best used for multi-hour agent runs and processing complex codebases.

Alibaba Qwen 3: Frontier tier with strong multilingual coverage.

This release shipped a frontier tier with strong multilingual coverage. It is best used for any deployment where Chinese-language quality matters.

Meta Llama 4: Native multimodal and long context, still open weights

Llama 4 shipped with native multimodal and long context capabilities, while remaining open weights. It's best used for multimodal applications that require end-to-end ownership.

xAI Grok 3: Trained on the Colossus cluster; competitive on reasoning benchmarks

Grok 3 was trained on the Colossus cluster, achieving competitive performance on reasoning benchmarks. It's best used for reasoning tasks, while remaining tightly tied to the X ecosystem.

Anthropic Claude 3.7 Sonnet: Extended thinking mode; toggleable deliberation for hard problems

Claude 3.7 Sonnet shipped with an extended thinking mode, offering toggleable deliberation for hard problems. It's best used for adaptive workloads, performing quickly on easy tasks and deeply on complex ones.

DeepSeek DeepSeek-R1: Open-weights reasoning model that rattled Silicon Valley.

This release shipped an open-weights reasoning model that rattled Silicon Valley. It is best used for math, code, and reasoning tasks with self-hosted control.

Google Gemini 2.0: Native agentic tool use and stronger multi-step reasoning

Gemini 2.0 shipped with native agentic tool use and stronger multi-step reasoning capabilities. It's best used for agentic browser tasks and Workspace copilots.

DeepSeek DeepSeek-V3: Frontier-class quality trained at a fraction of Western costs.

This release shipped frontier-class quality trained at a fraction of Western costs. It is best used for cost-sensitive assistants and as a reference for efficient training.

Alibaba Qwen 2.5 (+ Coder, Math): Family expansion with specialist coding and math variants.

This release shipped a family expansion with specialist coding and math variants. It is best used for coding copilots, math tutors, and structured extraction.

OpenAI o1 (reasoning): Chain-of-thought at the model level; trades latency for correctness

o1 (reasoning) shipped with chain-of-thought at the model level, trading latency for increased correctness. It's best used for math, hard coding, and formal reasoning tasks where waiting for an accurate answer is acceptable.

xAI Grok 2: Bigger jump in reasoning; image generation via Flux

Grok 2 delivered a bigger jump in reasoning capabilities and introduced image generation via Flux. It's best used for multimodal X integrations.

Mistral Mistral Large 2 / Nemo: Updated frontier tier plus a compact multilingual model

Mistral Large 2 / Nemo shipped as an updated frontier tier, alongside a compact multilingual model. It's best used for multilingual production LLMs on European infrastructure.

Meta Llama 3.1 (405B): First truly frontier-class open-weights model

Llama 3.1 (405B) was the first truly frontier-class open-weights model. It's best used for serious open-weight research and in regulated industries.

Anthropic Claude 3.5 Sonnet: Beat Claude 3 Opus on coding at Sonnet cost; the workhorse era

Claude 3.5 Sonnet surpassed Claude 3 Opus in coding performance while maintaining Sonnet's cost, ushering in the workhorse era. It's best used for production coding assistants and as agent backbones.

Alibaba Qwen 2: Substantial quality jump; competitive with Llama 3 tier.

This release shipped a substantial quality jump, competitive with the Llama 3 tier. It is best used for fine-tuning bases for enterprise Chinese/English apps.

DeepSeek DeepSeek-V2: MoE architecture with striking cost efficiency

DeepSeek-V2 features a Mixture-of-Experts (MoE) architecture, providing striking cost efficiency. It's best used for high-volume inference tasks where cost is the primary concern.

Mistral Codestral: Code-specialised model for autocomplete and refactor

Codestral is a code-specialized model designed for autocomplete and refactoring. It's best used for IDE assistants and self-hosted coding copilots.

Google Gemini 1.5 Flash: Cheap, fast tier of the 1.5 family for high-volume workloads

Gemini 1.5 Flash is a cheap, fast tier within the 1.5 family, designed for high-volume workloads. It's best used for real-time applications, feed rewrites, and batch classification tasks.

OpenAI GPT-4o: Native multimodal (text, vision, audio) with real-time voice

GPT-4o shipped with native multimodal capabilities across text, vision, and audio, including real-time voice. It's best used for voice interfaces, live translation, and mixed-media assistants.

Meta Llama 3: 8B and 70B tiers; approaches GPT-4 quality at self-hosted cost

Llama 3 was released in 8B and 70B tiers, approaching GPT-4 quality at a self-hosted cost. It's best used for on-premise enterprise LLMs and as a base for fine-tuning.

Anthropic Claude 3 (Haiku · Sonnet · Opus): Three tiers: Haiku fast/cheap, Sonnet balanced, Opus for hardest tasks

Claude 3 launched with three tiers: Haiku for fast and cheap operations, Sonnet for balanced performance, and Opus for the hardest tasks. It's best used for selecting the optimal point on the cost/quality curve for specific needs.

Mistral Mistral Large: Closed frontier tier; European sovereignty pitch

Mistral Large is a closed frontier tier model, positioned with a European sovereignty pitch. It's best used for EU-based enterprises with specific data-residency needs.

Google Gemini 1.5 Pro: 1M-token context — the long-context breakthrough

Gemini 1.5 Pro shipped with a 1M-token context window, representing a significant long-context breakthrough. It's best used for processing entire books, hours of video, or large codebases in a single prompt.

Alibaba Qwen 1.5: Multi-tier lineup from 0.5B to 72B, all open weights.

This release shipped a multi-tier lineup from 0.5B to 72B, all open weights. It is best used for multilingual production LLMs, especially APAC.

Mistral Mixtral 8x7B: Mixture-of-experts open weights; strong quality for the price

Mixtral 8x7B introduced mixture-of-experts open weights, offering strong quality for its price. It's best used for self-hosted assistants requiring GPT-3.5-class quality.

Google Gemini 1.0: Google's native multimodal frontier; Ultra, Pro, Nano tiers

Gemini 1.0 was Google's native multimodal frontier model, available in Ultra, Pro, and Nano tiers. It's best used for multimodal reasoning within the Google ecosystem.

xAI Grok 1: Elon Musk's first release; snarky persona, integrated with X

Grok 1 was Elon Musk's first release, featuring a snarky persona and integration with X. It's best used for X-native content and personality-driven chat interactions.

Anthropic Claude 2.1: 200K-token context; the first model that ate whole codebases in one shot

Claude 2.1 was released with 200K-token context, becoming the first model capable of processing entire codebases in one shot. It's best used for whole-repo code review and long-document Q&A.

Mistral Mistral 7B: Open-weights 7B model that punched above its size

Mistral 7B was an open-weights 7B model that performed remarkably well for its size. It's best used for edge, laptop, and cost-constrained inference scenarios.

Alibaba Qwen 1: Alibaba's first open-weights Qwen models.

This release shipped Alibaba's first open-weights Qwen models. It is best used for Chinese-language assistants and Alibaba Cloud deployments.

Meta Llama 2: First openly licensed model competitive with GPT-3.5

Llama 2 was the first openly licensed model that offered competitiveness with GPT-3.5. It's best used for self-hosted assistants and privacy-sensitive deployments.

Anthropic Claude 2: 100K-token context and stronger coding; the enterprise breakout

Claude 2 shipped with 100K-token context and stronger coding capabilities, marking its enterprise breakout. It's best used for legal, research, and analyst workflows involving long documents.

Anthropic Claude 1: Anthropic's first public model — long context, strong helpfulness training

Claude 1 was Anthropic's first public model, featuring long context and strong helpfulness training. It's best used for long-document summarization and safe assistant applications.

OpenAI GPT-4: First reliably 'smart' GPT — better reasoning, longer context, plus vision

GPT-4 was the first reliably 'smart' GPT, featuring better reasoning, longer context, and vision capabilities. It's best used for serious knowledge work, including analysis, coding, and structured extraction.

Meta LLaMA 1: Meta's first Llama — 'leaked' onto the open internet, changed everything

LLaMA 1 was Meta's first Llama model, which 'leaked' onto the open internet and significantly changed the Ai landscape. It's best used for kicking off the open-weights ecosystem.

Google Bard: Google's first ChatGPT response, initially built on LaMDA/PaLM

Bard was Google's initial response to ChatGPT, built on LaMDA/PaLM. It's best used for history books as it was quickly superseded by Gemini.

OpenAI GPT-3.5: Powers the original ChatGPT free tier — fast, cheap, conversational

GPT-3.5 powered the original ChatGPT free tier, offering fast, cheap, and conversational capabilities. It's best used for chat assistants, drafts, and summaries where cost is a higher priority than depth.

OpenAI GPT-3: 175B parameters; few-shot learning shocks the field

GPT-3 was released with 175 billion parameters, introducing few-shot learning that significantly impacted the field. It's best used for prompt-engineering-driven products, marking the era of 'just ask nicely'.

OpenAI GPT-2: 1.5B parameters, coherent long-form text, OpenAI briefly withheld weights

GPT-2 shipped with 1.5 billion parameters, producing coherent long-form text, leading OpenAI to briefly withhold its weights. It's best used for long-form generation demos and sparking early debates about Ai dangers.

OpenAI GPT-1: Unsupervised pre-training + fine-tuning for NLP tasks

This release proved out unsupervised pre-training combined with task-specific fine-tuning on NLP tasks. It's best used as a research foundation, demonstrating that transformers scale, rather than for production.

Can a Machine Think? Turing's Question That Started It All

In 1950, Alan Turing proposed the 'Imitation Game,' asking if a machine's behavior could define its intelligence. This fundamental question still frames every debate about whether Ai is truly intelligent today.

Dartmouth Workshop: The Moment 'Artificial Intelligence' Was Coined

In 1956, John McCarthy coined 'Artificial Intelligence' at a Dartmouth workshop, naming the nascent field. This event set an early tone of optimism that contributed to the first Ai winter years later.

The Perceptron: Ai's First Learning Machine and Its Early Setback

Frank Rosenblatt built the Perceptron in 1958, a hardware neural network that could learn. Its limitations, highlighted by Minsky in 1969, led to a significant funding freeze for neural approaches, showing Ai's stop-start history.

Expert Systems: Ai's First Commercial Wave Driven by Rules

The 1970s and 80s saw expert systems like MYCIN achieve impressive results with hand-written rules. While successful in narrow domains, their inability to generalize beyond these rules triggered the second Ai winter.

The Ai Winters: How Funding Crashes Reshaped the Field

Two major funding collapses in the 70s and late 80s forced Ai researchers to rethink their ambitions. These 'winters' still influence how practitioners approach product claims, making terms like 'AGI' career-limiting for many.

The Statistical Turn: When Machine Learning Overtook Symbolic Ai

The 1990s and 2000s saw statistical methods like SVMs and random forests dominate practical Ai applications. These pragmatic techniques still power a vast share of today's production Ai systems and competitive machine learning.

AlexNet Wins ImageNet: The Dawn of Deep Learning in 2012

In 2012, AlexNet dramatically reduced the ImageNet error rate using GPUs, marking the arrival of deep learning. This pivotal moment is the single most repeated inflection point, tracing back to every trillion-dollar Ai valuation today.

'Attention Is All You Need': The Transformer Paper of 2017

In 2017, Google Brain published the Transformer paper, introducing the architecture behind every modern Large Language Model. This paper is arguably the closest thing Ai has to a moon-landing moment, fundamentally changing the field.

The GPT Era Begins: Scaling Laws and Unexpected Capabilities

From 2018 to 2020, OpenAI's GPT series demonstrated the power of unsupervised pre-training and scaling laws. GPT-3's few-shot learning capabilities particularly surprised researchers, establishing the 'bigger is smarter' paradigm.

ChatGPT: Ai's iPhone Moment in 2022 Changes Everything

ChatGPT launched in November 2022, becoming the fastest-growing consumer product in history and making Ai a household name. This moment transformed Ai from a research topic into a boardroom priority, leading to the Yai Ai feed's existence.

Multi-Modal, Agentic, and Open: Ai Gets Loud From 2023-2025

Recent years brought multi-modal capabilities, agentic tool use, and open-weight models to the forefront of Ai. These advancements are making factory-floor use cases realistic, allowing Ai to be deployed beyond simple chat.

The Frontier Today: Claude 5, GPT-5, Gemini 2.5 in 2026

Today's frontier models like Claude 5, GPT-5, and Gemini 2.5 represent the state-of-the-art you're building upon. Every model in your Ai stack has a rich lineage, which you can now confidently name and understand.