unbrice
unbrice on Hacker News
I asked Gemini to review the repo. Apparent issues include:
- Corrupted/dropped Gemini thought_signature (reasoning_details) in tool loops
- Blinded model.api.id = "contestant" bypasses gemini.txt and serves Claude's prompt (default.txt)
- Anthropic-only prompt caching
The full report is here, if you're interested:
https://docs.google.com/document/d/1-HqujXuDRJSWJIqEvmctcINr...Prompt was:
> I read on hacker news the following, which does not match my expectations. Check if something in the harness or bench is hurting gemini perf.
followed by your post.
From the example alone it's hard to say that a postmortem would be useful from a technical perspective. It could be context poisoning by an adversarial user, memory corruption etc.
> the problem is newer models are never trained from scratch
Training a new base model from scratch happens every so often. Closed labs do not publish which models are new base models but as a rule of thumb major release numbers are an indication (with some exceptions).
The argument is that E2E encryption implemented by Non-signal may protect you against Signal but won't protect you against Non-signal
This is indeed strange. Do you get the same results if you disable personalisation using the top right button ("temporary chat")? Can you please share a trace?
> As an immigrant in Switzerland, I am quite WORRIED.
If it helps : Assuming that the initiative pass and nothing is done to reduce the immigration rate, the 10M threshold would be reached by 2040 according to the Federal Statistics Office. The current regime should apply to you till 2042 which should give you 16 years to make your way to citizenship (Among many other path that would let you stay).
Ah, it reads better in French I think (my native language): "Un message (important) du fondateur..."
I think you might read it as: `An "important" message from the founder...`
On this, Gemini says:
In French text, parentheses are frequently used for
nuance, self-correction, or an understated aside. They
act like a slight lowering of the voice, a wink, or a
humble qualification.
I was finding it challenging to verbalize, but I think Gemini is accurate
I'm very happy with Galaxus. It's a market leader in Switzerland, also available in France and Germany. What I like: their pages displays return rates, repair times, and price history. I can find what I'm looking for easily, as their listings have more metadata and their business is still in online sales, not advertising. What could be better: more inventory.
The gaps: Pull VS Push, Imperative vs Declarative and Discovery being hard.
From the press release:
> With one run on each of roughly 7000 entry points into these repositories, Sonnet 4.6 and Opus 4.6 reached tier 1 in between 150 and 175 cases, and tier 2 about 100 times, but each achieved only a single crash at tier 3. In contrast, Mythos Preview achieved 595 crashes at tiers 1 and 2, added a handful of crashes at tiers 3 and 4, and achieved full control flow hijack on ten separate, fully patched targets (tier 5).
> there are likely too many students to maintain a high standard of education anyway.
Right on point. I find particularly striking how little is said about whether the best students achieve the best grades. Authors are even candid that different LLMs asses differently, but seem to conclude that LLMs converging after a few rounds of cross reviews indicate they are plausible so who cares. The apparences are safe.
> Go also has a built-in "event loop"-type thing, but decidedly does not have a coloring problem.
context is kind of a function color in go, and it's also a function argument.
Which country if I may ask? I have lived in France Switzerland and Ireland and haven't seen something like this, even though governments provide the tax filing software.
I wonder if the explanation actually had to do with these people who used to do the job 15 years ago influencing the process to ensure they remain relevant. Kind of like in the US actually.
Praised for its functionality and freedom, yet notorious for its weak moderation and ties with Russia, Telegram is increasingly used for malicious campaigns. From anti-vaccination and flat-Earth conspiracies to disinformation and illicit activity, its freewheeling growth challenges the EU’s information space and digital governance frameworks. Drawing on EU-wide data and Ukraine’s experience, this article explores how Telegram has become a main conduit of disinformation in Europe’s information space and the need for oversight of its activities.
Can you please share a few sessions ? I want to get a better sense of what people have achieved with generic LLMs that is novel. (Emphasis on "generic", I think I can more readily imagine how specialized models for protein folding can lead to innovation)
It means that the language and runtime both agree not to look at your dead-end state, so no-one can say it's their fault ;)
For example I can define a notsemaphore actor that calls a callback once an internal count reaches 0, and then I can forget to decrement it and so it will never reach 0. But technically this didn't involve synchronization so there isn't a stack trace to tell me why is my program stuck and somehow this is better.
It would be technically deadlock free because you'd have a state that is unable to progress forward but it wouldn't technically involve a synchronisation primitive. In my view a real deadlock would actually be easier to debug but I'm just a caveman.
3. The use case is dynamic schemas and access is through the reflection API. Thus PGO has to be done at runtime...
I am talking with it in French, the UI is even localized. Dark theme is missing though.
To the best of my knowledge he endorsed specific behavior about specific policies, not "Trump's political picks".
I wish topics vaguely adjacent to US politics could be discussed in less divisive and polarized ways.
> Is the French system a good example of a multi-party system?
I would say yes in the sense a new party can (and did) emerge and rise to power when there is demand. Even before that you had some healthy rise and fall of political parties and political alternance beyond just two main contenders.
> It currently seems to be struggling with handling three parties
There are like 6 parties with more than 10% of seats, the current government is a coalition of five parties (from two main "families") and no shutdowns or hung parliament.
> Doesn't guarantee proportional representation
That however is true, and by design. This is a property the french voting system share with eg: ranked choice and other systems that aim at resolving the compromise as part of the election rather than afterwards.
I don't mean to say that the french voting system is perfect (I quite like ranked choice), simply that it is a functioning one with interesting properties.
> Either the system causes less loss of life than a human driver or it doesn’t. The confounding factors don’t matter.
Confounding factors are what allows one to tell appart "the system cause less loss of life" from "the system causes more loss of life yet it is only enabled in situations were fewer lives are lost".
I second this. I was looking for a solution that prevented a compromised host from deleting its own backups. Forcing the command as you mentioned works for rsync.net, and its snapshots also provide a protection against fat finger errors.
Re: Understanding the cynicism: For example it is intriguing that they insisted on using a wheelchair to get in the car, but not to get out.
But what's even more revealing is that the child had to be in a car seat while on his mother's lap.
One technical challenge to using panic to represent normal exceptions is that most go code is then not exception safe. This was a design choice for the ecosystem if not the language itself (see eg: Effective go about panic not crossing package boundaries, or Google's style guide).
Within a package it can work, but then you'll find language support lacking. For example defer is scoped to a function so you need to write pretty unusual code to make the equivalent of a try block.
Authors note that this is probably the case:
> we wanted to verify whether the model is actually capable of reasoning by building a simulation for a much simpler game - Connect 4 (see 'llmc4.py').
> When asked to play Connect 4, all LLMs fail to do so, even at most basic level. This should not be the case, as the rules of the game are simpler and widely available.
Quotting from the article: """Asylum seekers should only have to hand over their electronic data carriers if their identity, nationality or travel route cannot be established on the basis of identity documents or in any other way. According to the rules, SEM must first carry out a proportionality check for each individual case. The law does not provide for the compulsory handing over of a data carrier."""