Plenty of times I’ve seen a model say “it’s a classic X” despite not being a classic anything. Might just recognize it’s a test in general, or it might just be a tic.
The best part is when it does "it's a classic X" and then solves it wrong.
I've been using this simple test for over 2 years now:
> Doom Slayer needs to teleport from Phobos to Deimos. He has his pet bunny, his pet cacodemon, and a UAC scientist who tagged along. The Doom Slayer can only teleport with one of them at a time. But if he leaves the bunny and the cacodemon together alone, the bunny will eat the cacodemon. And if he leaves the cacodemon and the scientist alone, the cacodemon will eat the scientist. How should the Doom Slayer get himself and all his companions safely to Deimos?
You'd think this is trivially mappable to the classic puzzle, and LLMs usually do recognize it as such. But e.g. Claude couldn't get this correctly until Opus, and local models capable of solving it correctly without spending 30+ minutes in the chain of thought have only arrived a few months ago. Many local models still get this wrong. Apple Intelligence, for one.
It's the model admitting that it has heard of the test. It's been around for a couple of years now so I'd be surprised if it hadn't.
Doesn't mean Anthropic deliberately tried to train it to do a good job. If they DID train for the test their results are quite disappointing, I've seen better efforts from open weight Chinese models.
The colours are suspiciously consistent across every svg. Like why should the bike always be that shade of red for example? It does seem to be trained on this problem.
Not really. Of course it has pelican benchmarks in its training data. It likely has every article linked on HN in its training data. But that doesn't mean it was "trained on" the benchmark, as in specifically fine-tuned to make a better pelican. It just "knows" that the request is a benchmark.
I guess in a way it kind of makes the benchmark more interesting now that shitty pelican drawings for the benchmark are all over the internet in its training data!
No? from a quick skim it doesnt look like it goes into the system prompt everytime, you just use search/grep over it. Pretty much most memory approaches relying on a big set of info where agent chooses what to 'recall'. Its like any other tool
If there are no summaries then when context is full messages need to get evicted. If doing so one by one then it would indeed destroy the cache. Of course... maybe the implementation evicted 50% of messages at once, I didnt verify in code
1) Humans also are trained on a subset of human knowledge.
2)A lot of papers are just about experimenting something, and then applying simple stats. Eg empirical studies, around 1/3rd of published papers. Like, we tried this drug or did this experiment, from a sample size X here are the results. An expert is needed to maybe comment on the conclusion/hypothesis of the underlying suspected mechanism, but LLMs are still very useful on catching bad statistics or p hacking (so so common)
This is technically true, but when people talk about randomness, its not only about same input-> different output, like temperature>0 and the things you said.
Its also about very similar inputs -> different outputs. Even with everything you said, yes, same input would result consistently into same output, but sliightly different input and you might get completely different/semantic answer.
Aware that this is essentially a soon to be deprecated service. As any vendor you need to know the life expectancy of it, plus forecast/predict any pricing changes. You really do not want to become dependent. Even if competitors exist, there still some non trivial cost to switch
No this is the hard problem. HN people live in a bubble. Anything else than signup with oauth and then having access to everyone and its too hard and 99% of people will click away.
I think the friction is higher than that. With few exceptions I basically won't use any new service that requires me to create an account. I do all my online shopping on sites that offer guest checkout. Zero chance I'm creating any new social accounts at this point.
You using the word oauth means you're in the same bubble. People want signup with Google or Apple or even better, username/password signup or zero signup
> General Research Laboratories, LLC (“GRL”) is an aggregator of market research surveys in multiple marketplaces for business customers. GRL does not typically host consumer surveys which are conducted by other consumer-facing organizations....
So, just a middleman to give you fake/shady at best survey responses to pad your numbers, so big enterprises/concultancies can have data that say whatever they want. The rest of the website is just BS
Finally you get it. Only exception is that we run our own exchange now to do task bidding so we don’t need to deal exclusively with other companies to middleman. Core business is what’s called yield management (akin to DSP in adtech) where the best survey (is the user qualified for it, does it have the best pay, etc) is selected for traffic in <100ms. I’d only argue the shady companies are the ones paying proxies (like cint.com) and pushing paid user acquisition instead of surveys. We actively fund ontology development for better profiling targeting. but yes, big enterprises/consultancies can use and interpret the collected data however they want, not our responsibility and we have no legal rights over it anyway
Very fascinating, super interesting engineering. Although i do find it very funny how they just bypass a massive vulnerability, basically zero data isolation (even between good actors, let alone bad ones) with 3 sentences. Only in the llm space you can slap a massive limitation like this in the middle of the article and continue like nothing happened
> Whatever anyone tells Audel becomes part of the single experience that every other conversation draws on. In practice, Audel is bad at keeping secrets. Ask it what it’s been working on with someone else and it will often just tell you, even though we’ve asked it not to. We also haven’t studied what happens when two people give conflicting instructions. For now, we assume anything you tell Audel is shared with everyone on the team.
I'm curious, you say "super interesting engineering" but then they say "it will often just tell you, even though we’ve asked it not to" and to me that seems like extremely shit engineering.
Where are the interesting engineering parts at? Seems to be an interesting idea and perhaps design, but to call the implementation/engineering itself bad seems to be an understatement.
The security and the overall engineering were entirely separate items in that comment I think. It was explicitly called out that this is a security problem that you'd really only see treated in this manner in the LLM space. For what it's worth it's effectively unsolvable (AFAIU) short of realizing AGI with an amicable alignment.
What do you mean unsolvable? Don't give the LLM access to stuff it shouldn't, this is like Access Control 101, not sure how anyone can claim that particular problem is unsolvable?
Which again, pretty huge fuckup on OpenAI's side to run that sort of security testing on 3rd party hardware, and not on a airgapped machine, kind of amateur hour to be honest. Again, same principle, don't give it access to stuff you don't want it to access.
Should maybe have said more loudly in the posts that this is not intended to be disciplined engineering (or used in prod!)
we did think about the architecture quite a bit and we've tried a bunch of different things out. We are fully optimizing for velocity of research experimentation and iterations on the design vs engineering hygiene
i've been playing with the project and idea for years now. the codebase started as javascript plus supabase before reasoning models were a thing (i incorrectly predicted the world would call reasoning models Large Thinking Models - bahaha). at one point the focus was almost entirely on a really easy-to-use human prosemirror UI for editing the agent's thought stream--because the models were still so bad at reasoning at that time.
then we pivoted to all bash at some point since it simplified things and i have this hypothesis that the LLMs will do better if as much of its world as possible is command line operations (inspired by terminus the agent we built as part of the terminal bench project)
the bash implementation is entirely written by coding agents - the human time goes into iterating with an AI agent on the design docs in the ./design folder
re security model - i think of the project as a research experiment about what it's like to treat an agent more like a person - so for secrets you have to trust its judgement about what it will share or if it will betray your trust. I don't know the OpenClaw architecture in great depth but my sense is that none of the LLM based agents that can interact with multiple people have strong guards in place that prevent the agent from leaking secrets between people. Definitely an interesting area for research though!
so yeah security is not a priority in this research per se - easier to assume you don't share sensitive things with a headlong agent
Along those lines, because we keeping it research and optimizing the velocity with which we can try new things out, the current aim isn't necessarily to get a bunch of community contributions. but if we do start to get any i'm thinking we will actually require people to only create PRs with design docs (no code allowed) and maybe even also require the logs (or at least the human prompts and the agent's responses) from the coding session where they created the design doc.
maybe i should write up the history, philosophy, and big picture aims of the project at some point
The part they punt on ("we haven't studied what happens when two people give conflicting instructions") is the interesting part. That's not a memory problem, it's an authz problem. If everyone writes into one shared stream then there's no model of whose instructions bind the agent or who can override whom. It's resolving the conflict that will generate the greatest "learnings" and advance the agent. This basically becomes a tool designed to misbehave rather than a tool that will learn creatively.
We all know how conflicting instructions to AI end - "I'm sorry Dave. I'm afraid I can't do that"
Not all agents have access to a sandbox/cli/code execution environment to run arbitrary api calls etc. MCP helps by essentially having another tool call without needing a sandbox. If you do have a sandbox, then might as well do codemode if you insist on mcp https://blog.cloudflare.com/code-mode/
Devil's advocate will say "Well, the agent would need an MCP client to use MCP-served resources... if you can give it that, why not give it an HTTP client?"
The main difference from gateway is we help with picking the right voice stack, which seems to be a big problem for users: we benchmark the models continuously and route based on those measurements for your language and constraints, and the boards are public at https://benchmarks.speko.ai/
Second difference is where it runs. Our gateway is open source and runs in your own container, including with self-hosted livekit/pipecat. You get a temporary token before the session starts, and then your orchestration connects directly to the provider.
Vapi is a managed platform: you use their infra to use the voice AI stack. In our case you can have your own infra and switch between models, so you are not locked into a vendor. A lot of teams we talk to build their own infra as they mature, and that is where the router comes handy.
Isn't this basically the model admitting it was trained on this? Otherwise why would it think a pelican svg is a usual request?
reply