My focus was on finding a good text→URL-slug compression strategy. I used ChatGPT-5.2-Pro mainly to explore and compare different compression approaches and trade-offs.
If you’ve tried OpenAI’s Deep Research or similar tools, you’ll know they pull far more info than Wikipedia. But if you’re an expert, you’ll quickly spot errors since the breadth is huge but the depth and accuracy are only so-so.
For non-experts just exploring new topics, it’s still perfectly useful. Grokipedia probably uses a similar search, verify, summarize workflow, so it naturally inherits mistakes from the internet, which isn’t really an LLM problem.
Grok is just the first to make it public, and other AI companies could easily build their own synthetic data Wikipedias, and some probably already have.
Wikipedia’s coverage looks broad, but it still can’t keep up with how fast knowledge grows. And the gaps are even more severe in non-English versions of Wikipedia.
They should really back projects like Servo or Ladybird to cut down on their reliance on Chromium/Webkit. It’s the only way to truly support an open-source ecosystem and push forward the next generation of browser engine, even if it takes time.
Yeah but that requires time and effort. If they couldn't be bothered to launch on Windows and Linux when Chromium already has solid support for both, bringing up an immature browser engine was never going to happen.
That just shows they were in a rush to release the product rather than taking time to refine it. You can tell it’s quite rough and simple, with almost no distinctive visual style.
ChatGPT Atlas just packages several existing ChatGPT features into Chrome, for example, the basic Chat UI and an Agent mode. By turning the browser into the product, it gives them access to more user data, enabling more personalized recommendations. It also allows them to execute specific tasks once users are logged into certain services.
It seems this generation focuses more on GPU and AI acceleration rather than CPU. The M5 chip allows Apple Vision Pro to render 10% more pixels and operate at up to 120 Hz. It delivers up to four times the peak GPU compute performance compared with M4, provides 30% higher graphics performance, and offers 15% faster multithreaded CPU performance.
Bitcoin Core v30.0 now relays/mines txs with multiple OP_RETURN outputs and raises the default -datacarriersize to 100,000 bytes. This is a mempool policy change, not consensus.
The move has already gotten negative feedback from devs worried it invites more arbitrary “junk” data on-chain. Bitcoin Core’s stance is harm-reduction: OP_RETURN is prunable and cleaner than UTXO-stuffing workarounds that people have been using anyway.
Economically, more fee-paying bytes means miners are unlikely to reject it.
How does one figure this out? As in I want to know the comparable Deepseek or Llama equivalent (size-wise) and don't want to figure it out by trial and error.
Is it indeed the plan of Apple to eventually run such kind of models direcly inside a iPhone?
Or are the specs of any stateOfTheArt smartphone well below the minimum requirements of such "lightweight" models?
My focus was on finding a good text→URL-slug compression strategy. I used ChatGPT-5.2-Pro mainly to explore and compare different compression approaches and trade-offs.