writing on jonah vairon

running unsloth llms locally

Sat, 20 Sep 2025 00:00:00 +0100

After running local llms in emacs with strix halo, I realised unsloth provides documentation on running their quantised models, e.g.:

With the docs, I can set the recommended values for model temperature, context size, and sampling.

There’s also great guidance on offloading parts of a model to the CPU, particularly useful in the case of MoE models like Qwen3.

With only 64Gb of memory, this wouldn’t apply to me so much. But Qwen3-Next-80B-A3B-Instruct seems likely to change that.

local llms in emacs with strix halo

Thu, 21 Aug 2025 00:00:00 +0100

problem

I got an HP ZBook Ultra G1a for a new laptop. The main draw being the strix halo processor it comes with. My particular machine has the AMD Ryzen™ AI Max+ PRO 395 and 64Gb of unified RAM. The idea being to run some relatively large local LLMs at a usable speed.

But, while powerful on paper, the software stack isn’t quite there yet to take advantage of the strix halo hardware. The latest versions of ROCm, for example, don’t support the processor line and aren’t readily available on Fedora 42 (which works perfectly on the Zbook Ultra).

towards fully-automated citation screening

Tue, 19 Aug 2025 00:00:00 +0100

abstract

Citation screening is one of the most resource intensive stages of conducting a systematic review, making it ripe for automation. And while machine learning tools have been developped over the years to try and assist researchers during the screening stage, these have predominantly been semi-automated solutions relying on part of the screening process to still be performed manually by experts. The most promising attempts at fully-automating the process have been in the form of classifier models, with most recent approaches trying to incorporate large language models (LLMs) into the process. However, research as of yet has been largely exploratory, producing complex systems that rely on abnormal candidate citation features and varying degrees of expert input, that are optimised with manual prompts and benchmarked on small, ad-hoc datasets. This paper leverages advancements in language model (LM) prompting and AI engineering to create the the first fully-automated LLM-driven pipeline for biomedical citation screening. Designed with simplicity and minimisation of expert input in mind, the pipeline was automatically optimised and benchmarked on the CLEF2019 dataset, producing state-of-the-art results in key classification metrics such as: recall, F1, F3, and WSS, outperforming previous works in the field.

loading SSL certs and keys from memory for httpx

Sat, 16 Aug 2025 00:00:00 +0100

the problem

If you store your keys and certificates in a unified config file that’s loaded into your Python app on start, like we do at work, then you’ll face some trouble creating ssl contexts for your httpx clients.

The heart of the problem is you can’t load certificates and keys from in memory strings when creating SSL contexts – you have to pass file paths to be read from.

And while httpx has a workaround, in so far as a way to load certificates and keys from the environment through SSL_CERT_FILE and SSL_CERT_DIR, it doesn’t really solve our problem; we would have to break apart our application config.

the consistency of arithmetic is unknowable

Wed, 21 May 2025 00:00:00 +0100

introduction

We are concerned with answering whether it is possible for us to know that arithmetic is consistent, i.e. whether we can know the consistency of a theory of true arithmetic. I will argue that:

If we want our true theory of arithmetic to be axiomatizable and consistent.
And if we want provability in a formal system as our standard for knowledge concerning facts about arithmetic.
Then, for all intents and purposes, we cannot satisfiably know that arithmetic is consistent.

I will argue that because of Gödel’s second incompleteness theorem, we cannot construct a theory of true arithmetic that can satisfy both our conditions (i.e. 1. and 2.). Largely because Gödel’s theorem restricts the ability of consistent, axiomatizable theories of a certain strength from internally proving their own consistency by way of a canonical consistency sentence. Forcing us, as a result, to compromise either on the properties we would want of a theory of true arithmetic (i.e. 1.) or, on our standard for knowledge concerning the consistency of theories (i.e. 2.) – neither of which is desirable/satisfying.

minimum passports for worldwide visa free travel

Sat, 11 May 2024 00:00:00 +0100

the idea

Henley & Partners released their 2024 passport rankings earlier this year.

But I have more than one passport and I wanted to know the combined power of my passports, not just the power of a single nationality.

And it turns out that on the road to answering that question we can also figure out what the strongest pair and trio of passports are, just for fun.

We can even go so far as to try and answer the question: what’s the smallest number of passports you would need to travel to every destination in the world without a visa?

'rm -rf' sftp python script

Tue, 12 Mar 2024 00:00:00 +0000

problem

Why does SFTP not allow for a simple rm -rf?

It’s a rhetorical question because there’s probably a reason.

But that doesn’t help me, because I need to be able to clear out and remove directories when updating this blog over SSH File Transfer Protocol and doing so manually is tedious.

So the task is simple: write a script to clear out and remove a particular directory over SFTP à la rm -rf.

the best fibonacci implementation?

Sat, 09 Mar 2024 00:00:00 +0000

I’ve been meaning to write about this ever since I came across it in Abelson & Sussman’s wizard book.

I mean, Fibonacci is the classic example of a recursive function. And it’s often just implemented as:

def fib(n):
    if n == 0 or n == 1:
        return 1
    return fib(n-1) + fib(n-2)
print(fib(5))

And while this seems all well and good on the surface, it turns out that this is far from an efficient implementation of the sequence. In particular, we end up computing calls to fib that we’ve already computed. In other words, we perform a lot of redundant computation.

three ways to check for anagrams

Thu, 07 Mar 2024 00:00:00 +0100

problem

Given 2 strings, check that they are anagrams.

solutions

worst

Perhaps the most intuitive approach: look at each character in the first string and check it’s in the second string by iterating through its characters. If a character in the first string isn’t present in the first, return False. Otherwise, remove the matched character from the second string so it isn’t matched again going forward.

def are_anagrams(s1: str, s2: str) -> bool:
    # no point proceeding if the strings aren't even of the same length
    if len(s1) != len(s2):
        return False
    for ch1 in s1:
        in_s2 = False
        for i, ch2 in enumerate(s2):
            if ch1 == ch2:
                in_s2 = True
                break
        if in_s2:
            # remove matched char from s2
            s2 = s2[:i] + s2[i+1:]
        else:
            return False
    return True
x = 'fried'
y = 'fired'
z = 'fixed'
print(are_anagrams(x,y))
print(are_anagrams(x,z))

True
False

While straightforward, it’s the worst approach because its \(O(n^{2})\).

recreating nyt games' spelling bee

Wed, 27 Dec 2023 00:00:00 +0100

I’ve been spending enough time playing NYT’s Spelling Bee puzzle game to be compelled to reproduce it in Python; especially seeing how straightforward it seems to be.

All I need is a list of English words and the alphabet, no?

Without any clever workarounds the creation of a game just requires the following:

randomly select 7 letters from the alphabet, without replacement.
take the first letter as the center letter in the bee hive (hexagon grid).
get the set of English words that can be formed from our letters (most finicky part)
if that set includes a word containing all 7 letters (i.e. a pangram), then we have a valid spelling bee game!

random letters

Easy! Python has the random package:

causal theories of reference solve informative identity

Fri, 21 Apr 2023 00:00:00 +0100

introduction

We are concerned with the informative identity puzzle: explaining the substantive quality of non-trivial identity statements with coreferring names. In particular, how true sentences of the form “\(a\) is \(b\)”, for two different names \(a\) and \(b\), can be in any way informative if \(a\) and \(b\) share the same reference. I will argue that causal theories of reference, as Kripke conceived them, can do away with this problem while remaining reductive, by representing the informative content of such identity statements as an expression of a non-trivial identity relation between distinct causal chains that corefer. For it is possible to understand causal chains as different from causal networks, where a causal chain is a specific path through an object’s (referent) causal network, causally linking the use of a name to the object it was originally baptised to. Hence showing that there may exist multiple causal chains in the same causal network, whose relationship is not necessarily made obvious to a speaker using coreferring names.

a generalised evolutionary theory of knowledge

Fri, 21 Jan 2022 00:00:00 +0000

introduction

The concept of evolution originally arose in Biology, yet, it has managed to find its way into a particular branch of contemporary epistemic¹ thought known as Evolutionary Epistemology (EE) (Bradie and Harms 2020). Having breached its disciplinary boundary by finding application within the domain of Philosophy, evolution as a concept, has been elevated to the rank of superconcept, according to Wilson (Wilson 2010). And rightfully so, having proven valuable to “one of the greatest philosophers of science of the twentieth century” (Thornton 2021), Karl Popper, as evidenced by works such as “Objective Knowledge: an evolutionary approach” (Popper 2012).