PROJUKTI BRIEF
1 Anthropic discovered Claude has a hidden mental workspace that formed by itself during training.
New research from Anthropic found that Claude developed an internal structure called a J-space, a kind of mental scratchpad where the model works through concepts without writing them into its visible chain-of-thought. Nobody designed it. It emerged on its own during training and now plays a measurable role in Claude's higher-level reasoning. The finding changes how researchers think about what is actually happening inside large language models when they appear to be thinking. Watch how it works or explore a live J-space demo here.
2 Chinese labs released two competitive open-source models, including the first trillion-parameter coding model trained on Chinese chips.
Tencent released Hy3, a strong open-source model for its size, while Meituan fully open-sourced LongCat-2.0, the first trillion-parameter coding model reportedly trained on 50,000 Chinese chips rather than Nvidia GPUs. Both arrive as cheaper, capable alternatives to US models. More significantly, LongCat-2.0 is a direct demonstration that China's push to build AI infrastructure independent of American silicon is producing real results.
3 Nvidia's most powerful AI rack system has reportedly been delayed from 2027 to 2028.
The Kyber system, a cabinet that fuses 144 Rubin Ultra chips into a single computer purpose-built for training frontier models, has reportedly slipped a full year according to SemiAnalysis. Nvidia has denied the report. If the delay is real, it extends the window in which competitors have room to close the hardware gap, and adds further uncertainty to the timelines of labs that plan to train next-generation models on Kyber-class infrastructure.
PROJUKTI FORWARD
The real question. For all the noise about AI changing everything, the more useful question is a simple one: which companies put AI to work on something specific, measured what happened, and got a number back worth reporting? A recent analysis by Detailed tracked 30 companies that did exactly that. Three patterns show up across nearly all of them.
Search that converts. Used-car marketplace CarGurus built an AI search tool called CG Discover and confirmed on its February earnings call that the tool drove a 3.5x increase in website traffic and a 10x jump in attributable leads on a quarterly basis. Amazon and Walmart report similar conversion gains from their own on-site AI tools. The pattern is consistent enough across retailers that AI-powered search is no longer an experiment. It is a conversion layer.
Content at a pace humans cannot sustain. Duolingo and Coursera, two companies widely considered vulnerable to AI disruption, used AI to go on offense instead. Duolingo's co-founder Luis von Ahn said AI allowed the app to ship 10 times more courses than it could two years ago. Coursera now has more than 120,000 learners using its AI course-dubbing feature. In both cases the bet was straightforward: use AI to produce more content faster, and let distribution do the rest.
Personalization that scales. QuickBooks rebuilt its entire onboarding flow using AI to detect each new user's business type and generate a tailored welcome experience from the first screen. That kind of individualized treatment at scale was previously impossible without a large customer success team. The remaining case studies from Detailed's report follow a similar logic: find a workflow that requires human judgment at scale, hand it to AI, and measure the result.
PROJUKTI SNAPSHOT
Prompt that improves every session. One Redditor always opens with "What are you least confident about right now?" before starting any AI task, and swears by the results. The comment section responded with four other top-voted openers worth saving to your prompt library.
Creepiest screensaver on the internet. A former Google DeepMind engineer used Claude Fable 5 to build an infinitely scrolling VHS-style recreation of The Backrooms, the internet mythology of endless empty hallways. The technical execution is impressive. The result is genuinely unsettling in a way that is hard to stop watching.
AI superpowers, literally. People are using Gemini Omni's video editing capabilities to levitate cars, crush trees, and fake other impossible physical feats with convincing results. The comment sections are equal parts amazed and unsettled about what this means for video as evidence.
Jensen's jacket, at auction. Nvidia CEO Jensen Huang's signature leather jacket is going under the hammer at Sotheby's. The starting bid is considerably higher than you are probably guessing.
Token loophole, probably temporary. One developer found that converting dense text prompts into image format significantly cuts token usage and cost. Learn how it works here before the model providers close it.
How to set up agent loops. Anthropic's developer team published a detailed breakdown of how to get started with loops, covering what a loop is, the different types, and how to configure one from scratch. The clearest explanation of the concept published so far.
