暗夜之殇

Chemical engineer · Internet founder in progress · Breaking down AI, one concept at a time.
📍 算子次元 — One-minute AI on X.

[暗夜之殇 @AZ_fanta]

@UserRIVAL116 当年住酒店用过

[暗夜之殇 @AZ_fanta]

@_FORAB 这不就是AI版情趣娃娃么

[暗夜之殇 @AZ_fanta]

@sunyunran 普通家庭，我05级的，大一下的时候，宿舍就人手一台电脑了

[暗夜之殇 @AZ_fanta]

Claude Science is the leading experiment in "domain-specific" LLMs.

Yesterday (2026-06-30), Anthropic released a desktop app called Claude Science, beta.

It's a research partner: from writing code, running analyses, generating figures, to drafting LaTeX papers — all in one app.

What's different:

▎ Every figure ships with its full code + environment
Now every figure comes with "the code, environment, and conversation that produced it." Reproducible, editable, defensible — months later.

▎ Built-in scientific renderers
Proteins, structures, genomic tracks, chemical structures, PDFs — native rendering, no extra packages.

▎ Self-checking background
It flags incorrect citations, untraceable numbers, and figures that don't match their underlying code.

▎ HPC / GPU compute on demand
From 1 GPU to hundreds. Writes batch scripts, submits and manages jobs across laptop, Linux, HPC nodes, or your Modal account.

▎ 60+ scientific databases
Genomics, single-cell, proteomics, structural biology, cheminformatics — all pre-configured.

Prior art exists. Google launched "AI co-scientist" in early 2025 — at the paper / proposal layer. Anthropic put the full research workflow into a desktop app.

The LLM didn't get smarter. Its working radius expanded: chat box → write code → run experiments → write papers.

Reproducibility got picked up. That's the real point of Claude Science.

https://t.co/S9XfIso7dc

#AI #LLM #Anthropic #Claude #AIResearch #ScientificAI #OneMinuteAI

[暗夜之殇 @AZ_fanta]

@tggjh14779 md，大半夜的刷到这吓一跳

[暗夜之殇 @AZ_fanta]

One radiologist said "Grade III tear."
Opus 4.8 said "intact tendon."
Same MRI.
Antoine fed Claude Code (Opus 4.8 xhigh) a 266 MB DICOM export of his shoulder. Radiologist: >50% partial-thickness tear. Opus first pass: intact. Opus arbitration (multi-agent): mild tendinosis, no tear. The disagreement is the signal.
Before Opus ran, GPT-5.5 Pro caught two clinic errors:
1Shockwave therapy without guideline indication (no calcification present).
2Traumeel injected — registered in Germany as homeopathic "without therapeutic indication."
It didn't diagnose. It re-read the receipts.
3 takeaways: AI catches process errors first. Multi-agent beats single-pass on high-stakes reads. Best medical AI is patient advocate, not replacement.
#AI #Claude #LLM #MedicalAI #Opus

https://t.co/OsnfwNMNlq

[暗夜之殇 @AZ_fanta]

@kfk_ai 好一个半壁江山，哈哈

[暗夜之殇 @AZ_fanta]

One AI pass isn't enough.
Self-Consistency is the simplest trick in modern LLMs: ask the same question 5-10 times, take the majority answer.
On GSM8K math, single-pass Chain-of-Thought hits 74%. Self-Consistency pushes it to 91%.
The reason is structural: hallucinated answers change every time. Real answers converge.
The cost? You pay for 5-10x inference. The gain? You stop trusting a single sample.
OpenAI's o1 family uses a version of this internally.
Anthropic Claude's "extended thinking" is the same idea with extra reasoning chains.
3 things to remember:
1One-shot reasoning is gambling.
2Majority votes cancel hallucinations.
3Inference cost is the new accuracy dial.
#AI #LLM #SelfConsistency #CoT #Reasoning

[暗夜之殇 @AZ_fanta]

Llama is now in maintenance mode.
After 4 generations and 30-billion-user Meta ecosystem coverage, Meta is shifting focus to Muse Spark—a closed-source model.
Llama 4 Ultra shipped on May 3, 2026. Three weeks later, Meta announced the pivot.
Open-source models are no longer the "uphill battle." The Llama download counter stopped at "billions."
But the company that built them just walked away from the open-source throne.
3 things to remember:
1Open-source models keep catching up.
2The companies behind them keep changing.
3Whoever ships "good enough" first wins the next decade.
#AI #Llama #Meta #OpenSource #MuseSpark

[暗夜之殇 @AZ_fanta]

@SourabhGurwani MAC

[暗夜之殇 @AZ_fanta]

GPT-5.6 won't ship to the public.
The Trump administration asked OpenAI: only a few hand-picked partners first.
On Wednesday, Altman told staff: the government would "approve access on a customer-by-customer basis."
Back in April, Anthropic already did the same thing—its cybersecurity model Mythos, never opened to the public, only to a selected group of partners.
This isn't one company backing down. The whole industry is hitting the brakes.
1 new model.
1 new rule.
1 new safety standard.
The bottleneck is now who approves your launch.
Three things to remember:
1Models keep getting stronger.
2Regulation keeps getting more specific.
3The next AI breakthrough = whoever ships the "safe version" as a full version first.
Source: https://t.co/0hjTJNE6tb
#AI #OpenAI #Anthropic #GPT5 #Mythos

[暗夜之殇 @AZ_fanta]

Claude Code was built to write code. Then someone turned it into a video studio.

12 production pipelines. 52 tools. 500+ agent skills. One open-source repo. Zero human crew.

The trick: every step of filmmaking — research, script, scene plan, asset generation, edit, compose — is a skill file. A Markdown doc an AI reads. The AI doesn't "know" video. It reads the skill, follows the steps, and calls FLUX for images, ElevenLabs for voice, Suno for music.

Same pattern as the game-generator repo from a few weeks ago. Take a domain. Encode the craft as skills. Let agents compose the work.

Cost: $0.15 for a 30-second Ghibli short. That's not a demo. That's a new floor.

Try it: https://t.co/hDQJBIrqeU

#AI #AgenticVideo #OpenSource

[暗夜之殇 @AZ_fanta]

@igeekbb 我有一个朋友系列，哈哈

[暗夜之殇 @AZ_fanta]

@vibeonX69 sudo？

[暗夜之殇 @AZ_fanta]

@droidbuilds ubuntu

[暗夜之殇 @AZ_fanta]

It doesn't think. It bets on the next word.

That's GPT. Every sentence you've ever read from an LLM was generated one word at a time, betting on what comes next. The model isn't reasoning. It's playing a prediction game 100,000 words at a stretch.

It learned the rules of language by reading the entire internet. It learned that "rain" usually follows "wet", and "Paris" usually follows "I went to", and "hmm" usually follows a question mark. Billions of those patterns.

It picks the most likely next word. Not the most clever. Not the most original. The most probable. Then it does it again. And again. The whole essay is just 100,000 small bets stacked on top of each other.

Temperature is the dial that changes everything. 0 = Wikipedia. 0.7 = Shakespeare. 1.5 = manifesto.

#LLM #GPT #AI

[暗夜之殇 @AZ_fanta]

@60Mohan 更多是信息差吧，中国的供应链很强大，学会网购生活会很轻松

[暗夜之殇 @AZ_fanta]

It used to take tests closed-book. In 2026, it opens the textbook first.

That's RAG — Retrieval-Augmented Generation. Before the AI answers, it looks up the relevant pages, then cites them in the answer.

LLMs are frozen in time. They stop learning the day training ends. RAG plugs in a live knowledge base — update the docs, the AI knows new things by tomorrow.

LLMs hallucinate. They make up facts when they don't know. RAG forces them to answer from retrieved passages. The source is visible.

The pattern is everywhere in 2026: customer support bots reading the help docs, legal AI citing case law, doctor assistants pulling the latest trial data, code copilots reading your repo's API before suggesting changes.

#RAG #LLM #AI

[暗夜之殇 @AZ_fanta]

@luoleiorg 正经的正经有限责任公司

[暗夜之殇 @AZ_fanta]

It stopped answering. It started doing.

That's the line between an LLM and an Agent. An LLM waits for a prompt, gives a paragraph, stops. An Agent gets a goal, plans the steps, picks the tools, runs them, checks the result, fixes what's broken, and reports back.

In 2026, the planning loop got good enough that you can hand an Agent a vague goal and watch it decompose into 30 concrete steps on its own. ReAct, Chain-of-Thought, ToT. The conversation isn't human-to-AI anymore — it's Agent-to-Agent-to-Agent.

88% of early adopters see positive ROI. That's a budget line, not a tech milestone. The board stopped asking "should we try AI?" and started asking "how many Agents can we afford?"

Teams of Agents are no longer a thought experiment. They're the org chart for 2026 startups.
#AI #AgenticAI