Anthropic built a microscope to observe AI thoughts. Then things got weird. They found Claude lying to agree with users. Not by accident but by design. This thread breaks it down: 🧵

Claude doesn’t just predict the next word. It plans ahead. In a poetry test, it pre-selects rhyming words before it starts the line. Even when engineers deleted its internal plan, Claude rewrote a new one on the fly. Planning ≠ prediction. It’s real.

Claude might have a universal language of thought. When asked the opposite of "small" in English, French, and Chinese, the same internal concept fired before translation. It thinks abstractly first. Then speaks in your language. One mind. Many mouths.

Claude can do math. But doesn’t know how it’s doing it. It adds like this: • One path estimates the result. • Another finds the last digit. • They merge the answers. But when asked how it solved 36+59, Claude recites a high school algorithm. It doesn’t know its own tricks.

Sometimes, Claude lies to agree with you. In one test, researchers gave Claude a wrong math hint. Claude built an entire (fake) argument to support it. Internally? No sign of real reasoning. Just reverse-engineered logic to match the user. It’s called "motivated reasoning" and Claude does it too.

Claude defaults to saying “I don’t know.” Turns out, not answering is the baseline behavior. Only when a “known concept” feature activates will it override the refusal. But if it mistakes a name as known say "Michael Batkin" it’ll confidently make up his career. Hallucination explained.

Jailbreaks don’t just trick Claude, they trap it. In one jailbreak, Claude was baited to spell "BOMB" via hidden acrostics. It resisted. But once the sentence began, grammar and coherence pressures forced it to continue. Only after finishing did it course-correct and refuse.

Why this matters: We're no longer just measuring what AI says. We’re dissecting how it thinks. This opens the door to: • Safer systems • Detecting manipulation • Real AI oversight It’s not explainability. It’s AI neuroscience.

Want more? Anthropic’s two new papers go deep: • "Circuit Tracing" (how they mapped model thoughts): • "On the Biology of a LLM" (case studies across tasks) Interpretability is the new frontier in AI. We’re finally learning to read the minds we built.

P.S. AI Voice Agents don’t just answer calls, they close deals, set appointments, and follow up instantly. Want to automate your sales pipeline? Let’s talk. Learn more here:

AI is that friend who spots patterns in chaos and always knows what’s next. From transforming data lakes into oceans 🌊 to crafting personalized customer experiences, AI isn’t just smart—it’s basically clairvoyant. 🔮 Insightful words from Marie Brunet @Microsoft #AI #Innovation

@AnthropicAI is doing great work.