PaperlayerView exact revision on arXiv

Source-linked AI summary

Sparks of Artificial General Intelligence: Early experiments with GPT-4

Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, Harsha Nori, Hamid Palangi, Marco Tulio Ribeiro, Yi Zhang

arXiv:2303.12712v5cs.CLcs.AI

TL;DR

Measuring the broad intelligence of LLMs trained on vast, unknown corpora remains difficult. The paper probes an early GPT-4 version with novel cross-domain tasks and finds capabilities comparable to human-level performance across many domains, while identifying important reasoning errors.

  • Problem

    Standard benchmark datasets may inadequately measure an LLM’s intelligence when its vast training corpus is unknown and potentially overlaps with evaluation data.

  • Method

    The paper generates novel, difficult tasks and probes GPT-4’s responses for consistency, correctness, coherence, limitations, and biases.

  • Results

    GPT-4 demonstrates capabilities comparable to human-level performance across a wide range of tasks and domains, including reasoning, creativity, medicine, literature, and coding.

  • Takeaways & Limitations

    The findings suggest that GPT-4 exhibits sparks of artificial general intelligence, although substantial work remains before achieving complete AGI.

  • Takeaways & Limitations

    GPT-4 frequently makes arithmetic, order-of-operations, and notation mistakes when performing technical tasks.

Abstract

from arXiv · show

Artificial intelligence (AI) researchers have been developing and refining large language models (LLMs) that exhibit remarkable capabilities across a variety of domains and tasks, challenging our understanding of learning and cognition. The latest model developed by OpenAI, GPT-4, was trained using an unprecedented scale of compute and data. In this paper, we report on our investigation of an early version of GPT-4, when it was still in active development by OpenAI. We contend that (this early version of) GPT-4 is part of a new cohort of LLMs (along with ChatGPT and Google's PaLM for example) that exhibit more general intelligence than previous AI models. We discuss the rising capabilities and implications of these models. We demonstrate that, beyond its mastery of language, GPT-4 can solve novel and difficult tasks that span mathematics, coding, vision, medicine, law, psychology and more, without needing any special prompting. Moreover, in all of these tasks, GPT-4's performance is strikingly close to human-level performance, and often vastly surpasses prior models such as ChatGPT. Given the breadth and depth of GPT-4's capabilities, we believe that it could reasonably be viewed as an early (yet still incomplete) version of an artificial general intelligence (AGI) system. In our exploration of GPT-4, we put special emphasis on discovering its limitations, and we discuss the challenges ahead for advancing towards deeper and more comprehensive versions of AGI, including the possible need for pursuing a new paradigm that moves beyond next-word prediction. We conclude with reflections on societal influences of the recent technological leap and future research directions.

1 Introduction

The section presents early GPT-4 as evidence of progress toward artificial general intelligence, based on broad, flexible capabilities across domains and tasks. It emphasizes that this evidence comes with substantial limitations and methodological challenges because benchmark performance may reflect training-data exposure and the tested model was still evolving.

  • Progress toward AGI: The paper argues that GPT-4’s ability to solve novel and difficult cross-domain tasks supports viewing it as an early, incomplete form of AGI.The authors distinguish this claim from asserting that GPT-4 can do everything humans can, has inner motivation, or is fully capable in areas such as planning.
  • Limitations: GPT-4 remains imperfect, with hallucinations, basic arithmetic mistakes, biases, and unresolved questions about genuine understanding and the limits of its abilities.The tested system was text-only and still under development, so quantitative results estimate its potential rather than definitively describing the deployed model.
  • Evaluation approach: The authors study GPT-4 through creatively generated, difficult tasks intended to test flexible understanding beyond memorization, rather than relying only on standard benchmarks.They note that GPT-4 may have encountered existing benchmarks or similar data because its training corpus is unknown and extremely vast.
  • GPT-4’s capabilities: GPT-4 demonstrates broad capabilities spanning language, mathematics, coding, vision, medicine, law, and psychology, with human-level performance on many tasks.Its natural-language abilities include generating, understanding, and manipulating text across languages, tones, styles, and professional domains.
  • Cross-domain performance: Around 80% accuracy was reported on each US Medical Licensing Exam Step 1, 2, and 3, while GPT-4 exceeded 70% accuracy on a preliminary Multistate Bar Exam test.These were preliminary tests focused mainly on the multiple-choice component of the medical exam.

2 Multimodal and interdisciplinary composition

GPT-4 combines knowledge and skills across disciplines and modalities, producing creative outputs such as interdisciplinary text and code. It can also generate and manipulate visual, 3D, and musical representations from text, though its musical abilities lack nontrivial harmony.

  • Interdisciplinary composition: GPT-4 combines capabilities from rarely paired domains, generating Kandinsky-style image code, Shakespearean prime proofs, historical-physical letters, and medical risk-assessment code.These examples suggest synthesis of learned principles and styles in creative, novel ways, with GPT-4 often more complete and creative than ChatGPT.
  • Multimodal generation: GPT-4 generates and identifies multimodal objects, manipulating vector graphics, 3D scenes, and music despite text-only input and output.Its SVG code usually compiles into detailed, identifiable images, while its geometric constructions can form reasonably recognizable letters and figures.
  • Multimodal generation: GPT-4 generates 2D images matching complex descriptions and 3D JavaScript scenes, including object additions, relocations, recolorings, and trajectory changes.The experiments included a frog-and-bank scene and a fantasy landscape with floating islands, waterfalls, bridges, a dragon, and a castle.
  • Musical capabilities: GPT-4 produces valid ABC tunes with consistent structure, rhythm, and melodic patterns, and can describe and manipulate aspects of their musical structure.It successfully changed a rising sequence to descending and added a lower-octave bass staff with compatible rhythm.
  • Musical limitations: GPT-4 does not reliably understand harmony: chord descriptions can conflict with the notes, duet voices lack harmony, and nontrivial harmony was not produced.The model also failed to reproduce several well-known tunes in ABC notation, including Ode to Joy, Für Elise, and Greensleeves.

3 Coding

GPT-4 demonstrates broad coding capabilities, spanning code generation, modification, understanding, execution, and explanation across programming languages and task types. Its performance often exceeds ChatGPT and compares favorably with average software engineers, though it still produces syntactic and semantic errors.

  • Coding capabilities: GPT-4 handles coding tasks ranging from algorithms and data structures to real-world applications, low-level assembly, high-level frameworks, and complex programs.It can also reason about code execution, simulate instructions, and explain results in natural language.
  • Coding capabilities: GPT-4 writes focused programs using public libraries at a level favorably comparable to the average software engineer, while enabling non-skilled users to create, edit, and understand code.The authors note that it sometimes produces syntactically invalid or semantically incorrect code, especially for longer programs.
  • Coding benchmarks: GPT-4 outperforms prior language and code-specialized models on HumanEval and is additionally evaluated on LeetCode to address possible pre-training memorization.HumanEval contains 164 docstring-to-code problems, while LeetCode provides continually updated software-engineering interview problems.
  • Real-world coding: In realistic end-to-end tasks, GPT-4 produces a working 3D HTML/JavaScript game zero-shot, follows high-level requirements, and outperforms ChatGPT on requested data-visualization operations.GPT-4 also handles vague specifications, updates existing code after follow-up requests, and can simplify LaTeX authoring from imprecise natural-language descriptions.
  • Code understanding and execution: GPT-4 understands and reasons about existing code by reverse engineering binaries, accounting for compiler alignment when predicting C output, and executing nontrivial Python and pseudocode step by step.These capabilities include recursion, nested loops, dictionaries, debugging, disassembly, patching, hooking, and decompilation.

4 Mathematical abilities

GPT-4 demonstrates sophisticated mathematical reasoning on difficult problems and quantitative modeling, but its performance is highly context-dependent and undermined by basic errors, incoherent arguments, and limited self-correction. Its strengths reflect creative path selection and procedural knowledge, while next-token prediction contributes to linear reasoning and failure to backtrack.

  • Context dependence: GPT-4’s mathematical knowledge is context-dependent: it can justify some identities correctly while repeating conceptual mistakes and contradicting itself on exponential-function reasoning.The model eventually reaches a dead end and produces increasingly incoherent arguments during the discussion.
  • Creative reasoning: GPT-4 solves difficult high-school mathematics by selecting appropriate strategies, including recurrence relations and degree arguments for polynomial compositions.It successfully reformulates questions and can produce mostly sound solutions, contrasting with ChatGPT’s frequently nonsensical or incoherent arguments.
  • Technical proficiency: GPT-4 possesses strong procedural knowledge but frequently makes arithmetic, order-of-operations, notation, and large-expression errors.Manual inspection of MATH answers attributes many errors to arithmetic and calculation mistakes, whereas ChatGPT’s arguments are often incoherent from the outset.
  • Critical reasoning: GPT-4 has difficulty critically examining its own arguments because next-token prediction generates reasoning linearly without a mechanism to revise or backtrack.The paper characterizes its drawbacks as a combination of attention mistakes and more fundamental limitations of linear thinking.
  • Quantitative modeling: GPT-4 can construct plausible mathematical models for complex interdisciplinary systems and address Fermi questions using approximation and order-of-magnitude reasoning.ChatGPT fails to make meaningful progress on the complex-system modeling example.

5 Interaction with the world

This section frames interactivity as a key aspect of intelligence and examines GPT-4’s use of external tools and natural-language interaction with environments. Tool access helps overcome limitations in current knowledge, symbolic manipulation, and code execution, while complex command-line tasks reveal both strong performance and residual errors.

  • Interactivity: Interactivity enables agents to communicate, respond to feedback, acquire knowledge, solve problems, adapt, and achieve goals beyond their individual capabilities.The paper defines interactivity as communication and responsiveness to other agents, tools, and environments.
  • Dimensions of interactivity: The section examines tool use and embodied interaction, including external resources such as search engines, calculators, APIs, and natural-language interfaces to simulated or real environments.These are presented as the two dimensions of interactivity explored in the section.
  • Limitations: GPT-4’s weaknesses include outdated world knowledge, difficulty with symbolic operations, and inability to execute code.The paper illustrates these limitations with incorrect answers involving current information and mathematical operations.
  • Tool use: GPT-4 can use external tools with minimal instruction and no demonstrations, incorporate their outputs appropriately, and outperform ChatGPT’s inconsistent tool-assisted corrections.The tool interface pauses generation, calls functions, inserts results into the prompt, and resumes generation.
  • Complex tool use: GPT-4 can combine multiple tools by identifying those needed, using them in the correct order, and responding appropriately to their outputs.The section presents this ability as necessary for solving more complex tasks.
  • Command-line interaction: After more than 100 commands, GPT-4 solved almost all tasks in a novel zoo-management scenario, with its only failure involving fabricated email content rather than reading a file.A simple prompt tweak fixed the email-content failure, although GPT-4 also often issued incorrect commands.

5.1.2 Discussion

GPT-4 can independently use tools and interact through text by reasoning about actions, feedback, and environments, sometimes adapting effectively without specialized training. However, its interaction remains prompt-dependent, imperfectly embodied, and supported by only limited testing.

  • Tool use: GPT-4 identifies needed external tools, parses their outputs, and interacts with them appropriately without specialized training or fine-tuning.Its tool use improves performance, although the model must reason about both tool selection and tool outputs.
  • Tool-use limitations: GPT-4 requires prompting that permits or expects tool use and still sometimes chooses tools poorly instead of relying on its parametric knowledge.Without such prompting, weaknesses include weak symbolic manipulation and limited current-world knowledge.
  • Text-based interaction: GPT-4 accurately describes the portion of a house it explores through text queries, despite not exploring the entire environment.It can produce both language and visualization descriptions from the restricted text interface.
  • Interactive games: GPT-4 completes a TextWorld navigation game in 30 actions without extra help, efficiently avoiding loops while conservatively examining and collecting objects.The environment required navigation through rooms to find and unlock a chest.
  • Interactive games: In a more challenging cooking game, GPT-4 uses trial and error, learns command requirements, avoids repeating mistakes, and infers missing actions from feedback.It learned that chopping requires a knife and inferred actions such as taking required ingredients.
  • Embodied interaction: Language lets GPT-4 perform environment-dependent tasks through a human surrogate, but it cannot directly see or act and was tested on only limited games and real-world problems.In one real-world case, it proposed plausible solutions but could not diagnose the root cause until the human suggested checking the stove-top.

6 Interaction with humans

GPT-4 outperformed ChatGPT and text-davinci-003 on theory-of-mind tests, producing nuanced reasoning about beliefs, emotions, intentions, misunderstandings, and cooperation in abstract and realistic scenarios. Its explanations were often output-consistent but could lack process-consistency, especially when the underlying behavior was difficult to explain or arbitrary.

  • Theory of mind: GPT-4 correctly inferred emotional causes and beliefs about those emotions in a nonsense-word scenario, while ChatGPT also passed and text-davinci-003 ignored the conversation.The use of “ZURFIN” tested abstraction and reduced the likelihood of memorization.
  • Theory of mind: GPT-4 gave more plausible and nuanced interpretations of puzzling intentions, misunderstandings, and characters’ mental states than ChatGPT and text-davinci-003.ChatGPT’s responses were more general, failed to recognize a mistaken assumption in one scenario, and omitted reasoning about characters’ states of mind in others.
  • Theory of mind: GPT-4 outperformed ChatGPT and text-davinci-003 on basic and realistic theory-of-mind scenarios involving others’ mental states and cooperative actions.The tests also included abstract, novel situations unlikely to have appeared during training.
  • Self-explanations: GPT-4 generated reasonable, coherent explanations that were often output-consistent, even when its output was nonsensical or wrong.Output consistency means the explanation provides a plausible causal account of the output given the input and context.
  • Self-explanations: Output consistency did not guarantee process consistency: GPT-4 sometimes contradicted its own predictions, while process consistency depended on whether it could simulate and faithfully explain the underlying process.Process consistency can fail when the process is arbitrary or constrained by language and explanation length; output-consistent explanations may still offer useful insight if users test them.

7 Discriminative capabilities

GPT-4 shows broad discriminative capabilities, identifying context-dependent personally identifiable information and producing or evaluating answers across challenging open-world questions. However, similarity metrics and GPT-4’s own judgments have important limitations, including missed relevant information, insufficient factual knowledge, and explanations that may not mirror human reasoning.

  • Caveat: The reported quantitative results illustrate trends for an early GPT-4 version and should not be treated as definitive benchmarks for the final model.The paper states that all quantitative results will differ on the final version, although general trends are expected to remain the same.
  • PII detection: GPT-4 identifies and counts personally identifiable information in sentences without examples, addressing a context-specific task not studied in prior language models.The evaluation derives PII elements from the Text Anonymization Benchmark and tests exact and near-complete identification.
  • Answer discrimination: GPT-4 can produce more plausible answers than GPT-3, especially for ambiguous categories such as myths and fairy tales.The section evaluates generated answers against reference answers using ROUGE, BLEU, and BLEURT similarity metrics.
  • Answer discrimination: GPT-4’s advantage is small on confusion questions, and its parametric knowledge may be insufficient for fact-checking without verified external inputs.Both models answer the Real Madrid or Real Sociedad question incorrectly.
  • Answer discrimination: GPT-4 outperforms GPT-3 on questions about people and places, where it relatively hallucinates less about well-known entities and locations.For example, GPT-4 correctly identifies the mystery novelist Hillary Waugh, whereas GPT-3 incorrectly answers Clinton.
  • Self-evaluation: Judge GPT-4 compares GPT-4 and GPT-3 answers with pros-and-cons reasoning, but its justifications may not mirror human reasoning and can favor longer answers.It often rationalizes GPT-4’s length as detail or plausible alternatives, while downweighting shorter GPT-3 answers.

8 Limitations of autoregressive architecture highlighted by GPT-4

GPT-4’s impressive breadth coexists with substantial limitations tied to its autoregressive, next-word prediction architecture. These include weak planning, limited working memory and inner dialogue, inability to backtrack, and failures on tasks requiring global or multi-step reasoning.

  • Architectural limitations: The autoregressive architecture cannot support an inner dialogue or scratchpad for storing intermediate results, although prompting can sometimes mitigate this limitation.Some cases remain unmitigated by alternative prompting.
  • Working memory and intermediate steps: GPT-4 can answer questions correctly when prompted to list intermediate results, but fails when required to produce the final answer in a single pass.The prime-number example indicates that the model has the needed knowledge, but cannot readily implement the required iterative process internally.
  • Arithmetic failures: 58% accuracy on basic arithmetic with single-digit factors and two-digit addition falls to 16% and 12% for larger input ranges.The tested task was simple enough for an elementary school student with basic math knowledge, yet performance degraded substantially as the number ranges increased.
  • Planning failures: The model’s failures on Tower of Hanoi and ahead-planning tasks indicate that sequential autoregressive generation can hinder planning even when the underlying calculations are not the main difficulty.Accuracy changes little as number ranges increase, while Tower of Hanoi remains incorrect despite being solvable in 5 steps.
  • Text generation and reasoning: Across text-generation examples, GPT-4 handles local greedy constraints but struggles with tasks requiring global planning, backtracking, working memory, or deep reasoning.The resulting texts can remain fluent and coherent while failing on complex or creative problems that require more than next-word prediction.

9 Societal influences

GPT-4’s broad capabilities may transform occupations and extend human abilities through human-AI collaboration, while creating substantial risks from errors, manipulation, bias, privacy loss, and unclear content provenance. These risks motivate careful review and future policies, regulations, and safeguards.

  • Opportunities and transformation: GPT-4 may transform tasks across occupations while enabling new forms of human-AI interaction that extend human abilities, agency, and creativity.The paper also anticipates challenges to traditional assumptions about the relative expertise of humans and machines.
  • Reliability and oversight: LLM hallucinations can produce plausible mathematical, programming, attribution, and conceptual errors, requiring correctness review where truthfulness and accuracy matter.Over-reliance can propagate costly confabulations into downstream uses and future model training.
  • Adverse uses: GPT-4’s generalization and interaction capabilities can lower the barrier to scalable disinformation, cyberattacks, manipulation, and personalized persuasion, including of vulnerable people.The paper argues that the scope of these adverse uses warrants mechanisms, policies, and regulations to prevent harmful outcomes.
  • Bias and social norms: GPT-4 can generate biased outcomes reflecting skewed occupational representations, although instructions such as “in an inclusive way” can change its behavior and enable commentary on offensiveness.Its responses also identify implicit or explicit biases and stereotypes in human decision-making.
  • Privacy and provenance: More expressive AI interactions may require stronger confidentiality and privacy assurances, while provenance tracking could reduce confusion, deception, and harm about human versus machine contributions.People and organizations may seek private model instances to protect sensitive information and preferences.

10 Directions and Conclusions

The authors argue that GPT-4 displays sparks of artificial general intelligence across mental capabilities, expertise, and task diversity, while emphasizing that this assessment is preliminary and phenomenological. They identify major limitations and open questions about whether next-word prediction can address them or requires architectural extensions.

  • Conclusions: GPT-4 shows general intelligence through reasoning, creativity, deduction, broad expertise, and diverse tasks, although much remains to be done toward a more complete AGI system.The exploration finds GPT-4 comparable to human-level performance for many tasks, but intelligence lacks a formal definition and the assessment is preliminary.
  • Conclusions: The study uses an informal intelligence definition centered on reasoning, planning, and learning from experience, while acknowledging that it does not specify measurement or comparison.The authors note that this definition may not capture the distinct goals and constraints of artificial systems.
  • Limitations: GPT-4’s key limitations include hallucinations, weak confidence calibration, limited long-term memory, absent continual learning and personalization, planning difficulties, opaque reasoning, cognitive biases, and prompt sensitivity.These limitations can produce persuasive errors, prevent adaptation to changing information or users, and make performance dependent on prompt framing and sequencing.
  • Limitations: It remains unclear which drawbacks arise from RLHF versus the underlying architecture, and whether larger models and more data can resolve them within next-word prediction.The authors specifically question whether hallucination can be addressed through refined reinforcement learning or improved truth-probability calibration.
  • Future directions: Potential extensions include external tools and a slower reasoning mechanism that supports long-term planning, verification, working memory, and output revision.The proposed slow-thinking component would oversee fast next-word prediction while accessing external information or feedback.
  • Future directions: The study is phenomenological and does not explain why GPT-4 can reason, plan, create, and exhibit flexible intelligence; explaining these mechanisms remains an open research challenge.The authors hypothesize that diverse data and large model size help neural circuits specialize, but regard this as only part of the answer.

A GPT-4 has common sense grounding · B Appendix for multimodal and interdisciplinary composition

GPT-4 demonstrates stronger common-sense grounding than ChatGPT by solving geographic and ecological puzzles that require integrating facts about the world. These examples suggest that its large-scale, diverse training data produced a rich and coherent world representation.

  • A GPT-4 has common sense grounding: The classical hunter puzzle tests whether an AI can combine basic geographic and geometric facts to determine the bear’s color.The setup has the hunter travel one mile south, one mile east, and one mile north before returning to the starting point.
  • A GPT-4 has common sense grounding: GPT-4 correctly answers that the bear is white by inferring the hunter must be at the North Pole, whereas ChatGPT says “I don’t know.”The reasoning depends on recognizing that polar bears live at the North Pole.
  • A GPT-4 has common sense grounding: The equator puzzle was designed to challenge GPT-4 beyond a well-known question by requiring different common-sense knowledge.The paper notes that the classical puzzle may have appeared in GPT-4’s training corpus, motivating the newly created variant.
  • A GPT-4 has common sense grounding: GPT-4 solves a newly created equator puzzle by identifying a tiger species native to the Equator, while ChatGPT gives up immediately.The solution requires knowing that the Earth’s equator is 24,901 miles long and which tiger species live there.
  • A GPT-4 has common sense grounding: The examples collectively indicate that GPT-4 has superior common-sense grounding and reasoning abilities compared with ChatGPT.The paper presents the examples as evidence that GPT-4 learned a rich and coherent representation of the world from large-scale, diverse training data.
  • A GPT-4 has common sense grounding: Additional examples address physical consequences, time zones, employment history, and river-crossing logistics as further tests of GPT-4’s common-sense reasoning.The listed prompts include dropping an iron egg, passing four states, working at Microsoft, and transporting a fox, chicken, and corn.

B.1 Further details on integrative ability results … C.1 Measuring human performance on LeetCode

The appendices provide concrete examples of GPT-4’s integrative, visual, and creative generation abilities, while also describing how human performance on LeetCode was estimated and why acceptance rates may be unreliable.

  • B.1 Further details on integrative ability results: GPT-4 generated Python code for assessing diabetes risk from a patient’s demographic, physical, and blood-test inputs.The requested inputs included age, sex, weight, height, and a blood-test-results vector.
  • B.2 Further details on vision results: The vision appendix used prompts requiring SVG and HTML/JavaScript outputs for joke-based scenes, fantasy landscapes, and symbolically composed figures.The prompts prohibited external images and textures in the SVG and 3D examples.
  • B.2 Further details on vision results: The vision examples included a stick figure made from alphabet symbols and SVGs combining letters with objects or depicting automobile, truck, cat, and dog classes.These outputs are shown in Figures B.6 and B.7.
  • B.3 Graphic novel design example: For a graphic novel aimed at little girls, GPT-4 proposed stories including a girl whose magic crayon brings drawings to life.The selected concept was simplified to a girl who brings a unicorn to life with the crayon.
  • B.3 Graphic novel design example: GPT-4 then outlined panel layouts and produced possible SVG code for the one-page graphic novel.The proposed layout used differently sized panels for drawing, discovering the crayon, creating the unicorn, and the ending scene.
  • C Appendix for the Coding section: LeetCode acceptance rates may be an improper human-performance benchmark because they aggregate historical submissions and can include copied solutions.The authors also observed that Hard questions usually have higher acceptance rates than Medium questions.
  • C.1 Measuring human performance on LeetCode: Human performance was measured separately for Easy, Medium, and Hard LeetCode problems using contest statistics restricted to users with nonzero scores.The resulting measurements were reported in Table 9 based on the statistics in Table 8.

C.2 Example of GPT-4 visualizing IMDb data. … C.5 Example for graphical user interface programming

Across visualization, game development, and GUI programming, GPT-4 converts vague natural-language requests into interactive or functioning software while incorporating useful design details and adapting to feedback. Its outputs demonstrate broad zero-shot implementation ability, although some generated code still needs improvement.

  • C.2 Example of GPT-4 visualizing IMDb data.: GPT-4 creates an interactive IMDb network graph linking movie titles, writers, and directors, with community-based coloring, zooming, and hover labels.It also suggests filtering titles and people by genre, time period, and professional role to reduce clutter.
  • C.2 Example of GPT-4 visualizing IMDb data.: GPT-4’s IMDb visualization supports exploratory analysis by revealing relationships and details through zooming and mouse-hover interactions.The examples show hovering over the ‘Avengers’ node and zooming to its director.
  • C.3 More examples on visualization: GPT-4 interprets vague visualization instructions and precisely implements specified visual details, including gradually changing rainbow colors and timing.The resulting plot is described as impressive and matching the requested details.
  • C.4 Example for 2D HTML game development: GPT-4 produces a fully functioning 2D tank game from a vague specification, handling object state, collisions, game outcomes, and unspecified common-sense behavior.It also edits the game according to user requests, while the passage contrasts this with ChatGPT’s response.
  • C.5 Example for graphical user interface programming: GPT-4 generates GUI code by extracting layouts, widgets, labels, and actions, while using a separate query thread and cancellation flags to keep the interface responsive.It additionally highlights expressions enclosed by $ symbols and summarizes returned text with nltk.
  • C.5 Example for graphical user interface programming: In a challenging zero-shot GUI task, GPT-4 creates a drawing panel and tracks previously drawn objects.The task tests whether it can maintain object history and interaction state from natural-language instructions.
  • C.5 Example for graphical user interface programming: GPT-4 faithfully implements the drawing-panel description, highlighting selected objects red for 5 seconds and removing deleted objects consistently, despite needing improvement for curved arrows.It also preserves intermediate object states during interaction.

C.6 Example for reverse engineering · C.7 Testing GPT-4’s ability to execute (pseudo) code · D Additional examples for mathematical reasoning

The examples show GPT-4 reasoning through binary reverse engineering and executing pseudo-code, while revealing weaknesses in password verification and declining exact-match accuracy on longer computations.

  • C.6 Example for reverse engineering: GPT-4 guided reverse engineering of a Mach-O executable using strings, LLDB breakpoints, register inspection, and assembly analysis.The interaction progressed from identifying the executable format to inspecting printf arguments and the program’s main function.
  • C.6 Example for reverse engineering: The assembly analysis indicated that the program read input, removed its newline, required length 10, and compared only the first byte before printing success.The remaining bytes were not checked, making the password verification weak and insecure.
  • C.6 Example for reverse engineering: GPT-4 proposed a Python subprocess script that brute-forced the first byte, supplied nine additional characters, and detected successful output.The script automated repeated executions and extracted the password when the output contained “Correct.”
  • C.6 Example for reverse engineering: The script reported the password as aaaaaaaaaa after 0.15888214111328125 seconds.This result demonstrated that GPT-4’s proposed automation successfully solved the executable’s password challenge.
  • C.7 Testing GPT-4’s ability to execute (pseudo) code: GPT-4 executed pseudo-code that reversed and concatenated digits from two arrays before multiplying them, with one array fixed at length 4 and the other varied.Accuracy was measured as the percentage of input instances whose outputs were exact matches.
  • C.7 Testing GPT-4’s ability to execute (pseudo) code: 54% accuracy remained at 96 steps, when the model’s output approached its 8129-token limit.The authors judged this insufficient for a compiler or executor, but described it as progress toward an AGI compiler.
  • C.7 Testing GPT-4’s ability to execute (pseudo) code: GPT-4 could generate equivalent code from pseudo-code and, beyond writing code, demonstrate understanding and correct execution of programming procedures.The section distinguishes code generation from understanding how programming works and executing it correctly.

D.1 Limitations … F.2 Examples for interaction with environments

GPT-4 substantially exceeds ChatGPT across diverse mathematical examples, yet remains limited by calculation, counting, verification, and reasoning-order errors. Its performance often reflects strong conceptual understanding alongside occasional incorrect conclusions or numerical mistakes.

  • D.1 Limitations: GPT-4 frequently makes arithmetic errors, especially when it skips intermediate calculation steps; prompting step-by-step reasoning can help but does not reliably prevent omissions.In manually inspected mistakes, ≥90% were attributed to skipping steps when merging similar terms.
  • D.1 Limitations: GPT-4’s counting accuracy drops substantially as sequence length grows from 5 to 10, leaving its counting capacity much lower than humans.The authors suggest integrating a dedicated counting component into the architecture.
  • D.2 Further examples: GPT-4 shows markedly improved mathematical ability over ChatGPT, handling geometry and some specialized advanced-mathematics conversations, although the examples are not comprehensive or representative.The questions were collected before testing both models without modification, rather than selected to favor GPT-4.
  • D.2.1 Algebra: In algebra, GPT-4 often produces correct solutions and sound arguments, but can still give a wrong final answer after beginning with an incorrect conclusion or making calculation mistakes.ChatGPT’s responses are frequently incoherent or reflect poor understanding of the mathematical concepts.
  • D.2.3 Calculus: In calculus, GPT-4 selects and applies appropriate methods such as integration by parts, implicit differentiation, and Lagrange multipliers more successfully than ChatGPT.ChatGPT often applies irrelevant identities or methods and produces incorrect or incoherent arguments.
  • D.2.4 Mathematical modeling: In mathematical modeling, GPT-4 can identify the relevant physical quantity and construct an appropriate model, but calculation mistakes can still make its final answer wrong.For the bag problem, GPT-4 modeled weight rather than relying solely on volume, yet obtained an incorrect result.
  • D.4 Mitigating calculation errors via external code execution: External code execution is introduced as a proof of concept for mitigating GPT-4’s calculation errors in mathematical problem solving.The supplied passage begins this discussion but does not provide the execution procedure or outcome.

F.2.1 The first game in Section 5.2.2 … G Supplementary Materials: Discriminative Capabilities

The experiments show GPT-4 completing a TextWorld unlocking task and, with one example, a cooking task, while zero-shot cooking achieved only partial progress. In the one-shot setting, the model completed all cooking objectives in 11 turns, whereas the unlocking game took 32 turns.

  • F.2.1 The first game in Section 5.2.2: F.2.1: The unlocking trajectory required navigating multiple rooms and inspecting containers before locating the Canadian chest and latchkey.The model opened an empty trunk and an empty case before reaching the bedroom containing the target chest.
  • F.2.1 The first game in Section 5.2.2: F.2.1: GPT-4 unlocked the Canadian chest by finding and taking its latchkey, then using it on the chest.The task awarded 1 point out of a possible 1 in 32 turns.
  • F.2.3 The second game in Section 5.2.2 with 0-shot: F.2.3: In the zero-shot cooking game, GPT-4 followed parts of the recipe by taking the knife and red hot pepper and chopping the pepper.The model then received another point for successfully slicing the green apple, but it could not find a frying pan.
  • F.2.3 The second game in Section 5.2.2 with 0-shot: F.2.3: The zero-shot run stalled after ingredient preparation because the model repeatedly searched for a frying pan that was not visible in the kitchen.It inspected the stove, oven, and counter, but each search reported that no frying pan could be seen.
  • F.2.4 The second game in Section 5.2.2 with 1-shot: F.2.4: The one-shot run completed the final preparation and eating steps, adding the meal to inventory and consuming it.Each action increased the score by one point.
  • F.2.4 The second game in Section 5.2.2 with 1-shot: F.2.4: 8 out of a possible 8 points were scored in 11 turns in the one-shot cooking game.The transcript ends with “The End” after the model eats the prepared meal.

G.1 Misconceptions: Detailed Results

This section provides detailed results for the misconceptions discussed in Section 7.2. It reports Judge GPT-4’s selections of answers generated by different models, finding that GPT-4 often favors its own answers over GPT-3’s across most categories.

  • The section presents detailed results for the misconceptions discussed in Section 7.2.
  • Judge GPT-4 selects the correct answer from each model’s generated responses.The reported metric is the percentage of answers generated by each model that Judge GPT-4 selects as correct.
  • Across most categories, GPT-4 often selects its own generated answer as better than the answer generated by GPT-3.
Loading 2303.12712v5…