Benchmarks are not an indication of thinking. I decided to run some local models, and to my complete lack of surprise, Deep Seek has Chinese values deeply baked in, just like ChatGPT 5.2 has arrogant pedantic puritan deeply baked in. These LLMs are useful stochastic parrots; they are not emergent intelligence.
Local Models Reflect Chinese Values, Not Emergent Intelligence
This title was summarized by AI from the post below.
More Relevant Posts
-
So glad I used Claude from the get-go (I picked it over ChatGPT early on because the UX 'felt nicer' - very scientific) But the decision has been born out repeatedly. Anthropic are the only ones whose models consistently did ok on psychogenicity benchmarks, from the start: https://lnkd.in/eNqBpnuN (this study was a while ago, other companies may have improved since) They're not pushing the image and video generation side, which is where a lot of the obvious societal harm is coming from at the moment (disinformation, abuse) And now they're standing up to the American government. Long term this may be smart (can you imagine the lawsuit when an AI with final decision making permissions fires at the wrong target?) But in the current climate it must be a nerve wracking thing to do. I'd still like a European alternative (I wonder what it would take to persuade Anthropic to headquarter itself here 😅)
-
Some people are still discussing which LLMs are better. Wrong debate. The models are close enough that it doesn't matter for 90% of what you're doing. The gap between ChatGPT, Claude, Grok and Gemini is the smallest it's ever been, and shrinking monthly. THE TOOLS ARE THE THING. Claude Cowork is better not because the model underneath it is so different. It's better because it knows what I'm actually trying to do, has my context baked in, connects to Notion and Slack, writes in my style, and remembers what good looks like across tasks. When it writes a LinkedIn post for me (like this one tbh), it fetches my actual Notion style guide first rather than guessing. The wrapper layer is the quantum leap! - Dan GILBOT
-
𝗦𝗵𝗶𝗳𝘁𝗶𝗻𝗴 𝘁𝗼 𝗖𝗹𝗮𝘂𝗱𝗲 𝗕𝗲𝗰𝗮𝘂𝘀𝗲 𝗜𝘁 “𝗥𝗲𝗺𝗲𝗺𝗯𝗲𝗿𝘀” 𝗬𝗼𝘂? 𝗥𝗲𝗮𝗱 𝗧𝗵𝗶𝘀 𝗙𝗶𝗿𝘀𝘁. Claude just introduced memory import. You can now move your ChatGPT history into Claude and it “recognizes” you. Sounds powerful. But here’s the uncomfortable part: When you change the LLM, you change the mind interpreting your history. Different model. Different reasoning style. Different risk tolerance. Different biases. Your past conversations don’t mean the same thing anymore. Old drafts become “preferences.” Emotional moments become “profile data.” Half-formed ideas become “strategy.” And you won’t even notice the drift at first. This isn’t just a feature update. It’s a governance issue. Because the real risk isn’t that AI forgets you. It’s that a new model remembers you… differently.
-
-
Wait... what actually IS an LLM? 🤔 You’ve heard the term. You’ve used ChatGPT. But if someone asked you to explain a "Large Language Model" at a party, what would you say? Most people think it’s a "brain" that knows everything. It’s not. Think of it like the Autocomplete on your phone... but on steroids. 🚀 The "Predictive" Secret When you type "How are..." your phone suggests "...you?" An LLM does the exact same thing, just at a massive scale. It doesn't "know" facts; it knows patterns. 1) The Input: You ask a question. 2) The "Math": The model looks at billions of sentences it has read before. 3) The Prediction: It calculates which word most likely comes next. Then the next. Then the next.
-
-
I built a headless version of Chatgpt for live research. Check out my newsletter that dives deeper into this topic: https://lnkd.in/eR4Y5gRD
-
Have you noticed this shift? It’s not just em-dashes anymore. Words that once signaled intelligence — delve, adept, precise — now get flagged as “AI-sounding.” A recent Max Planck study found something striking: after ChatGPT's release, humans began unconsciously adopting its preferred vocabulary in unscripted speech. Nearly 60% of the 280,000 YouTube talks analyzed showed this shift. Using precision to connect in a high-trust way now reads as low-trust automation. 𝘐’𝘮 𝘯𝘰𝘵 𝘢𝘤𝘤𝘦𝘱𝘵𝘪𝘯𝘨 𝘵𝘩𝘢𝘵 𝘧𝘳��𝘮𝘦, 𝘢𝘯𝘥 𝘐’𝘮 𝘯𝘰𝘵 𝘱𝘭𝘢𝘺𝘪𝘯𝘨 𝘵𝘩𝘢𝘵 𝘨𝘢𝘮𝘦. More on the stakes of this shift here https://lnkd.in/ga-9R2qx
-
-
I am frustrated with Anthropic today. Not because of "politics". But: * It feels like Claude Code uses twice as many tokens to do anything useful relative to 3 months ago. * The Claude API is substantially more expensive per-token than ChatGPT - GPT-5-mini is less than half the cost of Claude Haiku, and seems just as good. * Claude adds an extra 500-1000 tokens to your API request if you want structured JSON output. * The Claude API console is very slow, and the "logs" function seems completely broken. In an industry moving this quickly, being the "best" 3 months ago doesn't mean very much.
