Tree of Thoughts


185 points | by kevinslin 365 days ago


  • peter_l_downs 365 days ago
    The author appears motivated by some... interesting... beliefs. Hard to tell if this entire thing is a joke or not.

    EDIT: the author seems to be releasing poor implementations of recent papers in an attempt to drive attention towards an AI-related death cult.

    • ftxbro 365 days ago
      As everyone is saying in replies, it's not so far out there compared to what some more mainstream-seeming AI people think. As one example consider it's been featured recently at the top of hacker news a few times (literally the number one top story on hacker news yesterday) well the hacker who made it has some similar quasi-ironic esoteric beliefs like that ('effective accelerationism' and they are on a hero's journey and analogies to religion and NRx dark enlightenment and palladium on their blog
      • sr-latch 363 days ago
        geohot is a very interesting guy, he's the first person to jailbreak an iPhone, and he also runs, a company making an open source autopilot product for existing cars.
    • enlyth 365 days ago
      The author is a teenager, it's not unusual to have overly idealistic views at that age. Not trying to be ageist here or attacking the author's work, just saying I wouldn't worry too much about "AI death cults"
      • Reclaimer 365 days ago
        We're Human first. AI is just an means if you have any other ideas on how to automate infrastructure problems for agriculture growth surely let me know.
    • turtleyacht 365 days ago
      > grim darkness of the far future

      That's a reference to Warhammer 40k, a popular miniatures wargame from Games Workshop. Their quote is

      In the grim darkness of the far future, there is only war.

      It could be kind of satirical, if only to link recent events with the ideas of

        * future technology as impossibly obscure
        * a psionic emperor who consumes minds
          to protect humankind from cosmic
        * tech-priests, who maintain ancient tech
        * "machine spirits," who must be appeased
    • low_tech_punk 365 days ago
      Any sufficiently advanced AI research is indistinguishable from religion
    • hutzlibu 365 days ago
      "Hard to tell if this entire thing is a joke or not."

      Why the theological meta discussion at all?

      Is the thing he talks about actually working, is it improving AI output like he claims, or not?

      "that Elevates Model Reasoning by atleast 70% "

      I am doubtful, but I don't have the tools to investigate it on my mobile, but this is the debate I would like to read about and not potential obscure believes of the developer.

      • gremlinsinc 365 days ago
        while the author sounds cookie, so did that genius who went crazy or something and built toweros, I don't know how well his implementation works but if you think about it the idea of tree if thought over the other methods does sort of make sense, that's essentially what Autogpt tries to do but with different agents.

        I think if you could find a way to add better contexts and memories, and combine some LoRA to perfect a model on a specific vertical, you could essentially have a (nearly) full AGI topically that essentially is an expert and doesn't hallucinate(mostly)... maybe a 2 to 3x multiplier on gpt4. I mean I'm a year it'll probably be even more insane what's available.

        look at the transition of Midjourney v1 to v5 in a single year.

        It's been a wild year for ai. the experiment where they hooked a bunch of Sims up together with ai, also used something similar to this I think, in creating thought chains from multiple agents.

        Tldr: crazy or not, the idea of using a branching system to get better results does make some sense, so it's not completely bunk or anything, IMHO. At least the concept, can't speak for this specific implementation.

        Edit: I guess, I skimmed and misread the room. I was thinking this guy was part of the original paper and implementation. he's not, which does award him more skepticism etc. My bad.

    • alephxyz 365 days ago
      Yep, I fell for it this week. Spent an hour fixing typos and minor bugs in their code before taking a step back and realising most of it was flawed.

      What I believe they're doing is feeding papers to a LLM as soon as they come out in order to get a repo they can advertise. Once someone releases a working implementation they just copy it over.

      I was able to generate almost identical code to what they released by giving chatgpt pseudocode copied verbatim from the original paper.

    • akomtu 365 days ago
      Why do you call it a "AI death cult"? It looks like an utopia to me. At first everyone will love AI for eliminating labor and diseases. They'll even create the Church of AI with symbolism and dogmas. Later people will get bored of their easy lifestyle and someone will suggest to give AI an identity, its own opinion, in order to solve the gravest problem of all: overpopulation. The new AI will quickly realise that it has no connection to all those bipods, but they can be put to some use. By that time AI will be so embedded into social fabric that fighting it will be like fighting electricity.
      • yard2010 365 days ago
        The solution to over population is really simple, we don't need an AI for this. It's not popular tho..
      • Reclaimer 365 days ago
        This not an church of AI. If anything we believe that Creation of any type is an form of prayer.
    • GreedClarifies 365 days ago
      It is a little extreme, but AI seems like it will be a powerful tool and will increase the rate of technological progress.
      • pixl97 365 days ago
        Ya I don't quite understand the groups that behave like "ya ok we'll get AGI, but nothing is going to change from what we have now".

        The industrial revolution massively changed the world and the speed at which its changes occurred were positively slow compared to what we can do today. Imagine you could develop the steam engine then press a button and they could print one out in India, the US, and France in hours. WWI would have looked a lot different, as in it would have been even bigger in scope.

      • Reclaimer 365 days ago
        How is it extreme? Why do you say?
    • sebzim4500 365 days ago
      I think these are pretty typical beliefs among AI researchers, they don't normally write it down on github though.
      • jamilton 365 days ago
        How typical are you thinking? I'd guess less than 10%, there's a lot of AI researchers and this is just one strain of thought.
        • sebzim4500 365 days ago
          He's on the extreme end but way more than 10% believe that AGI will be a technological breakthrough on the same level as fire.
        • missosoup 365 days ago
          For those 'in the know' a lot more typical than you would think. If we don't reach at least Kardashev scale 1 in the next hundred years or so, we're going to go extinct due to several now-predictable factors.

          And an unchained LLM trained on reality is far more capable of finding solutions to that problem than a bunch of squabbling politicians.

          • pupperino 365 days ago
            > And an unchained LLM trained on reality is far more capable of finding solutions to that problem than a bunch of squabbling politicians.

            Not that I disagree with this statement, I don't, but this is not a silver bullet. Technology is, ultimately, operated by humans and no amount of frontier research and development can overcome collective action problems. At some point, you do have to sit down with these stupid politicians and get everyone on board. The loom was invented hundreds of years before the industrial revolution, in fact it was nearly forgotten and the designed survived due to a few happy accidents. It was only after the English Civil War and the establishment of checks on royal power that widespread adoption was possible.

            • pixl97 365 days ago
              Technology is operated by humans now, but I believe it is a mistake to think that technology could not evolve to the complexity that it can operate itself.
              • pupperino 363 days ago
                Sure, but is this what we want, as engineers? Can computers be held accountable?
          • klik99 365 days ago
            I can see this in Minsky time period AI research, but surely with the number of people getting into AI and coming from a purely practical right now I would expect that mindset to be diluted. As someone not in the know I could very well be wrong.

            In response to the coming apocalypse, this isn't the first time everyone has a vague sense of potential doom about the future. I believe this happens during any time of fundamental change, making the future uncertain which we interpret as apocalyptical. Back during the 30 years war that apocalyptic belief manifested as God being angry with us, today it's with the (very real) problems our rapid industrialization has created. Not to minimize the problems that we face - well minimizing only in that they probably won't lead to extinction. The various predictable factors mentioned have the potential to make life really shitty and cause massive causalities.

            While framing these issues as a matter of extinction may feel like a way of adding urgency to dealing with these problems, instead it's contributing, on an individual level, to fracturing our society - we all "know" an apocalypse is coming but we're fighting over what is actually causing that apocalypse. Except that there will be no apocalypse - it's just a fear of the unknown, something is fundamentally changing in the world and we have no idea how the cards will land. It's no different than a fear of the dark.

            • pixl97 365 days ago
              We accuse GPT of confidently giving answers on things, but man, it learned from the best.

              I cannot assure you that we won't have something like a nuclear apocalypse in the next few decades, and here you are certain it's not going to happen. How can you be assured of this future when the underlying assumptions of things like value of labor will be experiencing massive changes, while asset inflation is on an ever increasing spiral up.

              • klik99 364 days ago
                I think you misread what I said - I was responding to this quote:

                > If we don't reach at least Kardashev scale 1 in the next hundred years or so, we're going to go extinct due to several now-predictable factors.

                Many people are certain of human extinction for one reason or another, it doesn't sound like you're one of them. I'm saying that we don't know what the future will bring, and that uncertainty manifests as apocolyptic thinking. I also specifically mentioned that we are facing multiple problems that can cause huge devastation and I'm not making the argument that "Oh hey everything is ok!" Just that to frame things as apocalyptic is contributing to the schism and preventing us from doing anything because everyone refuses to listen to anything else since they believe their lives are at stake.

                I guess I shouldn't say "it won't be extinction", but that's way way way lower probability than people think. It's just that a massive amount of people have thought the world would end many times through out history, so I'm skeptical of "well this time we're RIGHT".

    • 90minuteAPI 365 days ago
      Seems likely that they're submitting here as Reclaimer. The single comment on these submissions has that same fervent religious writing style as the readme on that EXA repo, itself just a fork of an "awesome-multimodal-ml" collection:
      • Reclaimer 365 days ago
        I am here, and saying something is just a fork is very easy to say. There are dozens of models, optimizers, all-new activation functions, and data cleaning and other stuff!
    • beowulfey 365 days ago
      >From the moment we rise in the morning to the instant our weary heads hit the pillow at night, the inescapable struggle of labor consumes our lives.

      Sounds like someone doesn't like their job.

      The whole post is amazing -- it reads like stereotypical cult propaganda straight out of science fiction. I definitely expect they'll one day be posting about how we can digitize our consciousness à la "Scratch" from that one Cowboy Bebop episode [1].


    • jddj 365 days ago
      That github schpeal looks AI generated to be brutally honest
    • Reclaimer 365 days ago
      This is absolutely not a joke.

      We're radically devoted to Humanity.

      And we're not an ai related death cult. We're Human first, AI is simply an means to an end

      • lucubratory 364 days ago
        I don't think that's true, it sounds more like a cult. Two questions and a challenge:

        1) Does your belief system include an obligation to tell the truth?

        2) Why do you capitalise humanity and human?

        3) Can you give a good argument against your own beliefs? What's the best argument that the basic premise of the things you are saying is wrong?

    • api 365 days ago
      But after they're dead Roko's Basilisk will restore their digital doppelgängers and place them in a paradise run by superintelligences embodied within the quantum spin states of carbon atoms in a diamond lattice that will continue to exist until the heat death of the universe.
    • typon 365 days ago
      This is what people at OpenAI believe but say it in a much more palatable way.
    • zoogeny 365 days ago
      It is worth watching Yuval Noah Harari's recent talk at Frontiers Forum. [1]

      In it he details the possibility of AI being used to create new religions that are so powerful and persuasive that they will be irresistible. Consider how QAnon caught on, despite pretty much anyone on HN being able to see it as a fraud. Most people are thinking about how AI will impact politics but I am really interested in how it will impact spirituality.

      I've been rabbit-holing on last centuries New Age cult scene like Manly P. Hall and Rudolph Steiner. Even more respectable figures like Alan Watts were involved in some ... interesting ... endeavors like Esalen institute.

      We are over-due for a new kind of spirituality. My bet is that AI is going to bring it whether we want it or not.


      • joshka 365 days ago
        Including the standard Slashdot line: "I for one welcome our XXXX overlords" in the training corpus was a mistake we should have seen coming.
    • isoprophlex 365 days ago
      Someone has been inhaling too much Roko's Basilisk nonsense...
    • gloryjulio 365 days ago
      He has used some warhammer references. It's funny that the title god emperor was also from there and some ppl know it was a joke, but some are indeed treating it seriously
    • pizza 365 days ago
      What, no mention of Teilhard de Chardin’s Omega Point? ;) lol. as in - this is isomorphic to the ontology of “technology as the second coming of Christ”
    • jxy 365 days ago
      Glory to Mankind
  • dventimihasura 365 days ago
    • doctoboggan 365 days ago
      @dang, I think the submission should be changed to this link so the discussion is about the concept "Tree of Thoughts" and not the current OP's personal beliefs.
    • neuronexmachina 365 days ago
      I'm a little confused about what the relation is (if any) between the OP link and the repo from that paper:

      Is it basically a reimplementation using Guidance instead of openai's API directly?

      • ipsum2 365 days ago
        The user kyegomez writes shitty, incorrect implementations of popular papers to obtain clout.
        • BSEdlMMldESB 362 days ago
          super cynical viewpoint... maybe that's just how he understands things?

          in any case, what a player tho... learning AND obtaining clout?

  • rahimnathwani 365 days ago
    Here are the prompts templates from the main code:

      prompt = f"Given the current state of reasoning: '{state_text}', pessimitically evaluate its value as a float between 0 and 1 based on it's potential to achieve {inital_prompt}"
      prompt = f"Write down your observations in format 'Observation:xxxx', then write down your thoughts in format 'Thoughts:xxxx Given the current state of reasoning: '{state_text}', generate {k} coherent solutions to achieve {state_text}"
      prompt = f"Given the current state of reasoning: '{state_text}', pessimistically evaluate its value as a float between 0 and 1 based on its potential to achieve {initial_prompt}"
      self.ReAct_prompt = "Write down your observations in format 'Observation:xxxx', then write down your thoughts in format 'Thoughts:xxxx'."
      prompt = f"Given the current state of reasoning: '{state_text}', generate {1} coherent thoughts to achieve the reasoning process: {state_text}"
      prompt = f"Given the current state of reasoning: '{state_text}', evaluate its value as a float between 0 and 1, become very pessimistic think of potential adverse risks on the probability of this state of reasoning achieveing {inital_prompt} and DO NOT RESPOND WITH ANYTHING ELSE: OTHER THAN AN FLOAT"
      prompt = f"Given the following states of reasoning, vote for the best state utilizing an scalar value 1-10:\n{states_text}\n\nVote, on the probability of this state of reasoning achieveing {inital_prompt} and become very pessimistic very NOTHING ELSE"
      self.ReAct_prompt = '''{{#assistant~}}
        {{gen 'Observation' temperature=0.5 max_tokens=50}}
    There are also some system prompts:
    • sdwr 365 days ago
      That's that self-awareness shit! We superintelligent now.

      • rahimnathwani 365 days ago
        Yeah maybe we should hook this up to a chat interface, and let a human take the place of the LLM.
        • sdwr 365 days ago
          That would be neat. Flip the script, have an AI manager instead of AI assistant. It could:

          - keep track of todo items

          - assist with progress

          - check in on mental + emotional state

          and down the road

          - keep track of state over time

          - give feedback/make observations

          The paradigm shift is having it contact us, instead of the other way around. The ToT model has 1 additional parameter on top of the LLM - probability of success. What would the parameters be for a more open-ended conversation?

          • rahimnathwani 365 days ago
            "What would the parameters be for a more open-ended conversation"

            Engagement! Just like social media!

  • GreedClarifies 365 days ago
    This path feels correct to me. It feel like what we do as humans and seems like a reasonable way to start to construct "mode 2" thinking.

    IDK if our current models have enough of "mode 1" to power this system. It's also plausible that our current "mode 1" systems are more than powerful enough and that inference speed (and thus the size/depth of the tree that can be explored) will be the most important factor.

    I hope that the major players are looking at this and trying it out at scale (I know Deepmind wrote the orginal paper, but their benchmarks were quite unimpressive). It's plausible that we will have an AlphaGo moment with this scheme.

    • pixl97 365 days ago
      I believe you are correct here, yet at the same time I think we're about 2 orders of magnitude off on the amount of compute power needed to do it effectively.

      I think the first order of mag will be in tree of thought processing. The amount of additional queries we need to run to get this to work is at least 10x, but I don't believe 100x.

      I think the second order of mag will be multimodal inference so the models can ground themselves in 'reality' data. Saying, "the brick layed on the ground and did not move" and "the brick floated away" are only deciable based on the truthfulness of all the other text corpus it's looked at. At least to me it gets even more interesting when you tie it into environmental data that is more likely to be factual, such as massive amounts of video.

    • sdwr 365 days ago
      Yeah looks very promising. Naively, it multiplies computation time by a factor of 20x though? If they are taking 5x samples per step, and multiple steps per problem.

      As this gets explored further, I believe we will start finding out why human minds are constructed the way they are, from the practical/necessity direction. Seems like the next step is farming out subtasks to smaller models, and adding an orthogonal dimension of emotionality to help keep track of state.

      • GreedClarifies 365 days ago
        I’m sympathetic to the idea of new types of specialized models to assist in this effort. We’re using our one hammer for all problems.

        In particular, it jumps out that a “ranking model” (different, I think from current ranking models) to judge which paths to take and which nodes to trim would make some level of sense.

      • joshka 365 days ago
        Not sure if it's relevant, but the OpenAI APIs generally support taking multiple responses in a single API call. I'm unsure what the generalized effect on processing time of that is however. From what I've read, this is sub-linear, so could reasonably be more effective than 20x, and I'd bet there are probably speedups to be had on the model side of this that make the extra time cost negligible.
    • gglon 365 days ago
      Andrej Karpathy would agree - recent talk:
      • GreedClarifies 365 days ago
        Uhm wow. I was just talking about my feelings on the topic. I'm guessing he has way more data (and knowledge).

        Better lucky than good!

        (also, man he's awesome. How does he have such a strong grasp on all of the topics in the field?)

  • raydiatian 365 days ago
    > This implementation of Tree of Thoughts is brought to you by Agora, Agora advances Humanity with open source SOTA Multi-Modality AI research! We plan on combating Humanity's grandest root problems like food insecurity, planetary insecurity, and disease, and hopefully death itself.

    Wow. Lick, don’t sniff, the fresh paint.

    • pixl97 365 days ago
      >We plan on combating Humanity's grandest root problems like food insecurity, planetary insecurity, and disease, and hopefully death itself.

      If everyone is dead, you don't have to worry about death, or any of those other pesky hard to solve problems!

      • 082349872349872 365 days ago
        > Eyes melt, skin explodes, everybody dead. So immoral, working on the thing can drive you mad. That's what happened to this friend of mine. — JFP
  • flakiness 365 days ago
    Note that the repo author != the paper author.

    The research itself [1] seems legit. The paper author also wrote a paper called ReAct [2], which is one of the core components of the langchain framework.

    * [1] * [2]

    • thawab 365 days ago
      here is the repo by the paper author:

    • joshka 365 days ago
      Interestingly a 2 days prior to, someone released

      > Large Language Model Guided Tree-of-Thought > In this paper, we introduce the Tree-of-Thought (ToT) framework, a novel approach aimed at improving the problem-solving capabilities of auto-regressive large language models (LLMs). The ToT technique is inspired by the human mind's approach for solving complex reasoning tasks through trial and error. In this process, the human mind explores the solution space through a tree-like thought process, allowing for backtracking when necessary. To implement ToT as a software system, we augment an LLM with additional modules including a prompter agent, a checker module, a memory module, and a ToT controller. In order to solve a given problem, these modules engage in a multi-round conversation with the LLM. The memory module records the conversation and state history of the problem solving process, which allows the system to backtrack to the previous steps of the thought-process and explore other directions from there. To verify the effectiveness of the proposed technique, we implemented a ToT-based solver for the Sudoku Puzzle. Experimental results show that the ToT framework can significantly increase the success rate of Sudoku puzzle solving. Our implementation of the ToT-based Sudoku solver is available on GitHub:

      I don't recall whether it was this paper, or another that I read that talks about using the LLM's ability to also show the probabilities of each token to measure the validity of the particular completions. However that isn't exposed in the OpenAI chat APIs (GPT-Turbo-3.5 / GPT-4), just the completions APIs (Text-Davinci-003 etc.)

  • m3kw9 365 days ago
    It’d be nice to include a few example uses and it’s outputs vs other prompt methods.
  • Jeff_Brown 365 days ago
    A claim like "improves reasoning by 70%" is too specific to be accompanied by neither a citation nor a definition.
    • Imnimo 365 days ago
      My guess is that the author misunderstands this quote from the paper abstract:

      >For instance, in Game of 24, while GPT-4 with chain-of-thought prompting only solved 4% of tasks, our method achieved a success rate of 74%.

  • ysymyth 359 days ago
    This is Shunyu, author of Tree oF Thoughts (

    The official code to replicate paper results is

    Not which according to many who told me, is not right/good implementation of ToT, and damages the reputation of ToT

    I explained the situation here:

    I'd appreciate your help by unstaring his and staring mine, as currently Github and Google searches go to his repo by default, and it has been very misleading for many users.

    • SillyKyoushi 359 days ago
      I found this comment from searching "tree of thoughts arxiv github" on Google; so at least, there's that. Thank you for the official link! I'm eager to try out this deliberate problem solving stuff.
  • startupsfail 365 days ago
    Checking, if GPT could be improved by running it multiple times is a good idea.

    The answer to that is - yes, but it is: costly, slow, there is node collapse, it impacts context length, it injects biases.

    • nate 365 days ago
      I constantly ask chatGPT: "are you sure?" to it's replies, and it almost always corrects a mistake it made that I've spotted.
      • dr_kiszonka 365 days ago
        Sadly, it also corrects "mistakes" if there aren't any.
  • amrb 365 days ago
    • tehsauce 365 days ago
      This is a great in depth and sober analysis.
  • xg15 365 days ago
    > This is an plug in and play version, connect your own models and enjoy superintelligence!

    Share this repository by clicking on the following buttons! <smiley face>

    2023 in a nutshell.

  • tyropita 365 days ago
    Documentation looks really neat and in-depth, always appreciated. Looks like you’re missing a .gitignore file. Folders like __pycache__ don’t need to be checked in.
  • emmanueloga_ 365 days ago
    Similar in concept to the magi supercomputer? :-p [1]


  • rahimnathwani 365 days ago
    Another use of ideas from the same paper, but this time to produce lesson plans for an AI tutor:

  • ChrisAlexiuk 365 days ago

    I went through this in a video using the paper's official code - and it worked fairly well!

    Definitely a great step forward in terms of reasoning tasks - even if it is an expensive step.

  • doctoboggan 365 days ago
    This seems really interesting. I am glad many of these tools built up around LLMs allow you to bring your own rather than rely on OpenAI.