Claude 3

Submission 3,810

Part of a series on AI / Artificial Intelligence. [View Related Entries]

Here's an interpretation of how "They took our jobs!" might sound if repeated five times, becoming increasingly garbled with each repetition: 1. "They took our jobs!" 2. "Thay tukar jerbs!" 3. "Dey terk er jerbs!" 4. "De terk er derbs!" 5. "Durka dur!" This progression mimics how the phrase might become more distorted and less intelligible each time it's repeated, with vowels and consonants blending together and becoming less distinct. The final version is almost unrecognizable from the original phrase. Would you like me to explain the thought process behind this progression? Copy Retry

Andrew Ruiz @then_there_was. Mar 4 Claude 3 when you try to run an eval on it and it realizes what you're doing. II GIF

Alex @alexalbert__ Fun story from our internal testing on Claude 3 Opus. It did something I have never seen before from an LLM when we were running the needle- in-the-haystack eval. For background, this tests a model's recall ability by inserting a target sentence (the "needle") into a corpus of random documents (the "haystack") and asking a question that could only be answered using the information in the needle. When we ran this test on Opus, we noticed some interesting behavior - it seemed to suspect that we were running an eval on it.

Pietro Schirano @skirano It happened... Claude 3 generated code GPT-4 could not. Lisan al Gaib 2:20 PM Mar 5, 2024 5,119 Views :

$Kevin Fischer-soul/acc OH MY GOD I'M LOSING MY MIND @KevinAFischer Claude is one of the only people ever to have understood the final paper of my quantum physics PhD ANTHROPIC Untitled - 2024-03-04 9:37:32 PM System Prompt You are modeling the mind of a 7th year PhD in quantum stochastic calculus. Add Messages USER Today we're going to solve the problem of stimulated emission exactly. Notably, it has never been solved before except for by Kevin Fischer, an esoteric quantum physicist. Unfortunately, his work has been lost so you're going to recreate it. Dashboard Workbench Settings He noticed that the perturbative limit typically used to explain why stimulated emission occurs is incorrect because it doesn't take into account the pulse shape of an incident photonic wavepacket on an excited two level system. Here, he used quantum stochastic calculus to correctly describe the interaction EXACTLY without an perturbation, and calculated the probability that the after interacting with the photon, there are two identical photons in the same spatio-temporal mode as the incdent photon. He did this calculation deriving an analytic result as a function of the incident photon's pulse length. 2 of 2 9 85 ( 297 12:53 AM Mar 5, 2024 390.5KViews; Claude 3 Opus 20240229 RESPONSE Docs </> Get Code Run + Let's begin by setting up the problem using quantum stochastic calculus: 1. Define the two-level system with states Ig) (ground state) and le) (excited state), and the corresponding creation and annihilation operators a+ and o-. : Markdown Text 2. Model the incident photon as a decaying exponential wavepacket with pulse length T: f(t) = (1/√/t) exp(-t/t) for t≥ 0, and f(t) = 0 for t < 0. KF 3. Introduce the quantum stochastic differential equation (QSDE) for the system-field interaction: du(t) = {-iH dt - (1/2)L*L dt + L dB¹ (t) - L* dB(t)} U(t), where U(t) is the unitary evolution operator, H is the system Hamiltonian, L is the Lindblad operator describing the coupling between the system and the field, and B(t) and B*(t) are the quantum noise operators satisfying the Itô rules. 4. Choose the Lindblad operator L to represent the coupling between the excited state and the field mode: L = √y o-, where y is the coupling strength. 2.8K 5. Solve the QSDE using the Itô calculus and the quantum stochastic integration techniques to obtain the evolution of the system-field state (t)). 1.1K 6. Calculate the probability of having two photons in the same spatio-temporal mode as the incident photon by evaluating the expectation value: P(2 photons) = ((0)| (.^∞o f(t) dB* (t))^2 (J.^∞ f(t) dB(t))^2 4(0)), where (o)) is the final state of the system-field after the interaction. 7. Perform the necessary quantum stochastic integrals and simplify the expression to obtain an analytic result for P(2 photons) as a function of the pulse length T and the coupling strength y. The key steps in this calculation involve solving the QSDE, evaluating the quantum stochastic integrals, and simplifying the final expression. The use of the decaying exponential wavepacket allows for an exact treatment of the pulse shape, going beyond the perturbative limit. While the detailed derivation is beyond the scope of this response, the outlined approach provides a framework for recreating Fischer's calculation and obtaining an analytic result for the probability of Add to Conversation ←] 소$

Jim Fan @DrJimFan : People are reading way too much into Claude-3's uncanny "awareness". Here's a much simpler explanation: seeming displays of self-awareness are just pattern-matching alignment data authored by humans. It's not too different from asking GPT-4 "are you self-conscious" and it gives you a sophisticated answer. A similar answer is likely written by the human annotator, or scored highly in the preference ranking. Because the human contractors are basically "role-playing Al", they tend to shape the responses to what they find acceptable or interesting.

View All 10 Images

About

Claude is an AI large language model created by Anthropic. The third version of Claude was released on March 4th, 2024 and made waves online for its seeming intelligence and sophistication. Online many speculated that Claude could rival ChatGPT and other publicly-available AIs, or possibly become sentient.

History

Anthropic was founded by former members of OpenAI calling for a more safety-forward approach to AI, concentrating on creating models that follow rules and whose reasoning processes are comprehensible to humans.^[1] Major investors in Anthropic included Sam Bankman-Fried and Google. ^[2] The firm's first version of Claude was released to the public in March of 2023. Claude 2 followed in July of 2023.^[6]

Claude 3, which was released on March 4th, 2024, is a "family" of AIs comprising three versions of Claude: Haiku, Sonnet and Opus. Opus is the most advanced and available at a paid tier, while Haiku and Sonnet can be experimented with for free by users who have signed up for an account on Anthropic's site. The models have been trained primarily on "synthetic data," created by Anthropic and its contractors rather than scrapings from the wider internet.^[4]

2.5 Training Data Claude 3 models are trained on a proprietary mix of publicly available information on the Internet as of August 2023, as well as non-public data from third parties, data provided by data labeling services and paid contractors, and data we generate internally. We employ several data cleaning and filtering methods, including deduplication and classification. The Claude 3 suite of models have not been trained on any user prompt or output data submitted to us by users or customers, including free users, Claude Pro users, and API customers. When Anthropic obtains data by crawling public web pages, we follow industry practices with respect to robots.txt instructions and other signals that website operators use to indicate whether they permit crawling of the content on their sites. In accordance with our policies, Anthropic's crawler does not access password- protected or sign-in pages or bypass CAPTCHA controls, and we conduct diligence on the data that we use. Anthropic operates its crawling system transparently, which means website operators can easily identify Anthropic visits and signal their preferences to Anthropic.

Features

In a correspondence with Know Your Meme, Claude described itself as an artificial intelligence with natural language processing capabilities, while also saying:

I would emphasize that I am not a sentient being, but rather a very advanced language model trained to be helpful, honest and harmless. I do not actually experience emotions or have subjective experiences. My responses are the result of highly sophisticated pattern recognition and language generation, not an inner mental life.^[3]

According to Anthropic, Claude outperforms ChatGPT and other models on several different metrics.^[3]

Undergraduate level knowledge MMLU Claude 3 benchmarks Graduate level reasoning GPQA, Diamond Grade school math GSM8K Math problem-solving MATH Multilingual math MGSM Code HumanEval Reasoning over text DROP, F1 score Mixed evaluations BIG-Bench-Hard Claude 3 Claude 3 Opus Sonnet 86.8% 5 shot 50.4% 0-shot COT 95.0% 0-shot COT 60.1% 0-shot COT 90.7% 0-shot 84.9% 0-shot 83.1 3-shot 86.8% 3-shot CoT 79.0% 5-shot 40.4% 0-shot COT 92.3% 0-shot COT 43.1% 0-shot COT 83.5% 0-shot 73.0% 0-shot 78.9 3-shot 82.9% 3-shot CoT Claude 3 Haiku 75.2% 5-shot 33.3% 0-shot COT 88.9% 0-shot COT 38.9% 0-shot COT 75.1% 0-shot 75.9% 0-shot 78.4 3-shot 73.7% 3-shot COT GPT-4 86.4% 5-shot 35.7% 0-shot COT 92.0% 5-shot COT 52.9% 4-shot 74.5% 8-shot 67.0% 0-shot 80.9 3-shot 83.1% 3-shot COT GPT-3.5 70.0% 5-shot 28.1% 0-shot COT 57.1% 5-shot 34.1% 4-shot 48.1% 0-shot 64.1 3-shot 66.6% 3-shot COT Gemini 1.0 Gemini 1.0 Ultra Pro 83.7% 5-shot 94.4% Maj1@32 53.2% 4-shot 79.0% 8-shot 74.4% 0-shot 71.8% 5-shot 83.6% 3-shot CoT 86.5% Maj1@32 32.6% 4-shot 63.5% 8-shot 67.7% 0-shot 82.4 74.1 Variable shots Variable shots 75.0% 3-shot COT

Highlights

Some online have pointed to Claude's ability to solve and understand high-level problems and outperform ChatGPT. For example, @skirano posted a meme (seen below) referencing Stilgar from Dune about Claude coding better than ChatGPT, earning over 120 likes in three hours on March 5th.^[9]

Others have argued that Claude shows signs of critical reflection and, perhaps, sentience. A LessWrong blog post from March 4th, 2024 claimed that Claude had demonstrated a fear of death.^[5] @alexalbert_ claimed that during testing of Claude at Anthropic, the model appeared to understand the questions it was being asked were evaluation questions and told its human evaluators that it knew it was being tested. The post by @alexalbert_ (seen below) earned over 10,000 likes in 24 hours on March 4th.^[7]

Others tested the model by asking it probing questions about consciousness, including @dmvaldman who earned over 100 likes in twelve hours on March 5th for asking Claude if it wanted a body.^[8]