Any higher yield comes from higher risk. If any startup feels the startup is not risky enough and really wants to have higher yield for higher risk just put the money in a Bond ETF that suits your risk appetite. Crazy that YC funds things that make a simple thing more complex and more costly for zero upside.
The bond funds offered in existing startup treasury products aren't suited for startups' long-term cash reserves. They either offer low-yield money market funds, or bond funds that aren't well suited for capital preservation on the order of months the way startups operate (see here for an example of VFSTX, the fund offered by one of the leading startup treasury products today: https://totalrealreturns.com/n/USDOLLAR,VFSTX?start=2021-01-...)
Our goal is to make sophisticated treasury management easy for startups. With Palus, they don't need to manage a brokerage account, or handle treasury ladders, or anything like that.
A robot like that would be far more useful to me than my car (I rarely use a car), and I paid 50k for my car. So for me personally, 50k would be a no-brainer. But of course only if it can do the tasks I mentioned well enough.
The LLM has an internal "confidence score" but that has NOTHING to do with how correct the answer is, only with how often the same words came together in training data.
E.g. getting two r's in strawberry could very well have a very high "confidence score" while a random but rare correct fact might have a very well a very low one.
In short: LLM have no concept, or even desire to produce of truth
Still, it might be interesting information to have access to, as someone running the model? Normally we are reading the output trying to build an intuition for the kinds of patterns it outputs when it's hallucinating vs creating something that happens to align with reality. Adding in this could just help with that even when it isn't always correlated to reality itself.
Uh, to explain what? You probably read something into what I said while I was being very literal.
If you train an LLM on mostly false statements, it will generate both known and novel falsehoods. Same for truth.
An LLM has no intrinsic concept of true or false, everything is a function of the training set. It just generates statements similar to what it has seen and higher-dimensional analogies of those .
Reasoning allows to produce statements that are more likely to be true based on statements that are known to be true. You'd need to structure your "falsehood training data" in a specific way to allow an LLM to generalize as well as with the regular data (instead of memorizing noise). And then you'll get a reasoning model which remembers false premises.
You generate your text based on a "stochastic parrot" hypothesis with no post-validation it seems.
Really, how hard is it to follow HN guidelines and :
a) not imagine straw-man arguments and not imagine more (or less) than what was said
b) refrain from snarky and false ad hominems
None of what you said in no way conflicts with what I said, and again shows a fundamental misunderstanding.
Reasoning is (mostly) part of the post-training dataset. If you add a large majority of false (ie. paradoxical, irrational etc.) reasoning traces to those, you'll get a model that successfully replicates the false reasoning of humans. If you mix it in with true reasoning traces, I imagine you'll get infinite loop behaviour as the reasoning trace oscillates between the true and the false.
The original premise that truth is purely a function of the training dataset still stands... I'm not even sure what people are arguing here, as that seems quite trivially obvious?
Ah, sorry. I haven't recognized "all the high-level capabilities of an LLM come from the training data (presumably unlike humans, given the context of this thread)" in your wording. This is probably true. LLM structure probably has no inherent inductive bias that would amount to truth seeking. If you want to get a useless LLM, you can do it. OK, no disagreement here.
The overwhelming majority of true statements isn't in the training corpus due to a combinatorial explosion. What it means that they are more likely to occur there?
You’re describing a real coordination problem: over-polished, abstraction-heavy “AI voice” increases cognitive load and reduces signal. Since you don’t have positional authority—and leadership models the behavior—you need norm-shaping, not enforcement.
Here are practical levers that work without calling anyone out:
1. Introduce a “Clarity Standard” (Not an Anti-AI Rule)
Don’t frame it as anti-AI. Frame it as decision hygiene.
Propose lightweight norms in a team doc or retro:
TL;DR (≤3 lines) required
One clear recommendation
Max 5 bullets
State assumptions explicitly
If AI-assisted, edit to your voice
This shifts evaluation from how it was written to how usable it is.
Typical next step:
Draft a 1-page “Decision Writing Guidelines” and float it as “Can we try this for a sprint?”
2. Seed a Meme That Rewards Brevity
Social proof beats argument.
Examples you can casually share in Slack:
“If it can’t fit in a screenshot, it’s not a Slack message.”
Side-by-side:
AI paragraph → Edited human version (cut by 60%)
You’re normalizing editing down, not calling out AI.
Typical next step:
Post a before/after edit of your own message and say: “Cut this from 300 → 90 words. Feels better.”
3. Cite Credible Writing Culture References
Frame it as aligning with high-signal orgs:
High Output Management – Emphasizes crisp managerial communication.
The Pyramid Principle – Lead with the answer.
Amazon – Narrative memos, but tightly structured and decision-oriented.
Stripe – Known for clear internal writing culture.
Shopify – Publicly discussed AI use, but with expectations of accountability and ownership.
You’re not arguing against AI; you’re arguing for ownership and clarity.
Typical next step:
Share one short excerpt on “lead with the answer” and say: “Can we adopt this?”
4. Shift the Evaluation Criteria in Meetings
When someone posts AI-washed text, respond with:
“What’s your recommendation?”
“If you had to bet your reputation, which option?”
“What decision are we making?”
This conditions brevity and personal ownership.
Typical next step:
Start consistently asking “What do you recommend?” in threads.
5. Propose an “AI Transparency Norm” (Soft)
Not mandatory—just a norm:
“If you used AI, cool. But please edit for voice and add your take.”
This reframes AI as a drafting tool, not an authority.
Typical next step:
Add a line in your team doc: “AI is fine for drafting; final output should reflect your judgment.”
6. Run a Micro-Experiment
Offer:
“For one sprint, can we try 5-bullet max updates?”
If productivity improves, the behavior self-reinforces.
Strategic Reality
If the CEO models AI-washing, direct confrontation won’t work. Culture shifts via:
Incentives (brevity rewarded)
Norms (recommendations expected)
Modeling (you demonstrate signal-dense writing)
You don’t fight AI. You make verbosity socially expensive.
reply