Max Caldwell
LightNews LightNews
banner
growth.wtf
Max Caldwell
@growth.wtf
390 followers 460 following 520 posts
Bootstrapping a software company. In my free time I try to do things with data and statistics, or post random thoughts here.
Posts Replies Media Videos
growth.wtf
0/3 + "you need help"
January 23, 2025 at 3:47 AM
growth.wtf
As expected
January 23, 2025 at 3:46 AM
growth.wtf
"reasoning"
January 23, 2025 at 3:45 AM
growth.wtf
????
January 23, 2025 at 3:44 AM
growth.wtf
It's crazy to me just how consistently wrong the google search llm is.

Is this like, a quantized gemini 1 that can't even see the search results it is summarizing? That's the only explanation for why doesn't it seem to get even context from featured snippets that they should have cached already.
January 1, 2025 at 11:17 PM
growth.wtf
:/
December 12, 2024 at 4:10 AM
growth.wtf
Sure, here's an example:
huggingface.co/datasets/Ple...

I am no expert but it seems legit?
The composition of the dataset adheres to the criteria for public domain works in the EU and, consequently, all Berne-countries for EU authors: any publication whose author is dead for more than 70 years. Additionally, the initial consolidation of public domain status for cultural heritage operates in the EU under the 2019 Copyright Directive (art. 14). As of June 2024, to limit rights verification, we have retained exclusively titles published prior to 1884. The corpus will be expanded at a later stage to encompass late 19th century and early 20th century publications, after checking for public domain validity.
December 6, 2024 at 12:11 AM
growth.wtf
Something I've always been curious about with social networks is how much different actions lead to an audience. At scale, it would be interesting to learn how large accounts got large, but I'm also curious what gets people to follow. Is seeing a good post more influential? Or follow-backs? Well...
Alt text: Two posterior distribution plots displayed side by side. The left plot shows "Posterior ATE of 1 Follow on Following" with a mean of 0.4021 and 94% HDI ranging from -10.25 to 10.23, displayed as a bell curve centered near zero. The right plot shows "Posterior ATE of 1 Post on Following" with a mean of 2.759 and 94% HDI ranging from -6.404 to 12, also displayed as a bell curve. Both plots use a blue line to show the distribution and include x-axes ranging from approximately -30 to 40 and -20 to 30 respectively.
December 5, 2024 at 8:49 PM
growth.wtf
From 2022 - tsunami danger zones in the bay area per the mercury news: www.mercurynews.com/2022/10/12/m...
December 5, 2024 at 7:22 PM
growth.wtf
@alpindale.bsky.social Getting there, not quite getting the model injection code right yet.

I have to start my actual work now but when or if finish this I'll open source.
December 5, 2024 at 3:59 PM
growth.wtf
OK, got a little DND game running with github.com/ag2ai/ag2 -- my review is that it's a neat set of concepts but very difficult to use and quite buggy. and I was hard-pressed to think of an actual use case that would fit into this.
December 5, 2024 at 8:09 AM
growth.wtf
Hey @alpindale.bsky.social , I saw you're working on the ruliad project? Can you help explain how to get the model to follow scripted directions? I'm trying to write an ollama template to try to get it to use tools, or follow scripted tasks of any order.
December 5, 2024 at 6:50 AM
growth.wtf
deepthoughts-8b-tools is confused, but it's got the spirit?
December 5, 2024 at 6:17 AM
growth.wtf
Unfortunately it's just not trained for this. Even if you get it working with Ollama tool calling, the model is providing a total hallucination about half the time. and is unable to provide responses in standard tool calling format.
December 5, 2024 at 5:47 AM
growth.wtf
Can I fix it?
December 5, 2024 at 4:46 AM
growth.wtf
Tell me you're overthinking it without telling me you're overthinking it. Full reasoning chain in the alt text.
Hello The user has provided a simple greeting 'Hello'. This is not a complex problem, but rather a basic interaction. I need to understand the context and determine the appropriate response." }, { "step": 2, "type": "data_gathering", "thought": "I have all the necessary data in the user's input: a single word 'Hello'. There's no additional information or context provided." }, { "step": 3, "type": "analysis", "thought": "The greeting 'Hello' is a common English phrase used to greet someone. It's a polite and casual way to initiate a conversation." }, { "step": 4, "type": "hypothesis_formation", "thought": "Given that this is likely the start of a conversation, the most appropriate response would be to acknowledge the greeting and reciprocate with another friendly greeting." }, { "step": 5, "type": "solution_proposal", "thought": "I should respond with 'Hi!' as it's a simple, friendly greeting that matches the tone of 'Hello'." }, { "step": 6, "type": "alternative_consideration", "thought": "Other possible responses could include 'Hey there!', 'Welcome!', or just 'Hi' without an exclamation mark. However, 'Hi!' remains the most straightforward and widely applicable option." }, { "step": 7, "type": "evaluation", "thought": "The proposed solution of 'Hi!' is appropriate because it's brief, friendly, and doesn't assume any additional context that isn't provided in the original message." }, { "step": 8, "type": "decision_making", "thought": "Based on the analysis and evaluation, I will proceed with the response 'Hi!' as it best fits the given scenario." }, { "step": 9, "type": "verification", "thought": "Verifying the reasoning chain: The problem was correctly identified, the context analyzed appropriately, and the solution proposed aligns well with the expected behavior for a simple greeting exchange." },
December 5, 2024 at 4:21 AM
growth.wtf
it's happening
code: cd Documents Documents uv init test_deepthought_ag2
December 5, 2024 at 4:17 AM
growth.wtf
20% worse than o1 and claude-latest at reasoning evals, but better than Qwen-72b, which is interesting?
December 5, 2024 at 3:03 AM
growth.wtf
Update, tailwind4 totally kicks ass. This is all you need to create new utility classes usable in code... normal css and variables!
December 4, 2024 at 2:22 AM
growth.wtf
skill issue
warning text stating: "mccabe: Cyclomatic complexity too high: 19 (threshold 15)"
December 3, 2024 at 4:32 AM
growth.wtf
My cat has gotten in the habit of coming to yell at me until I let him sit on my chest around this time. Funny wat to be forced to take a work break. Not allowed to say no to the cat.
Close up of an orange cat looking into camera
December 3, 2024 at 2:27 AM
growth.wtf
Do you find 300 characters limiting? Would you like to express yourself in way more than a single post can contain? Use the alt text!

It can contain almost 6x as many characters before you hit the size limits.

Click the alt text to see Canterbury Tales. the first 2,000 characters
Whan that Aprille with his shoures soote, The droghte of March hath perced to the roote, And bathed every veyne in swich licóur Of which vertú engendred is the flour; Whan Zephirus eek with his swete breeth Inspired hath in every holt and heeth The tendre croppes, and the yonge sonne Hath in the Ram his halfe cours y-ronne, And smale foweles maken melodye, That slepen al the nyght with open ye, So priketh hem Natúre in hir corages, Thanne longen folk to goon on pilgrimages, And palmeres for to seken straunge strondes, To ferne halwes, kowthe in sondry londes; And specially, from every shires ende Of Engelond, to Caunterbury they wende, The hooly blisful martir for to seke, That hem hath holpen whan that they were seeke. Bifil that in that seson on a day, In Southwerk at the Tabard as I lay, Redy to wenden on my pilgrymage To Caunterbury with ful devout corage, At nyght were come into that hostelrye Wel nyne and twenty in a compaignye Of sondry folk, by áventure y-falle In felaweshipe, and pilgrimes were they alle, That toward Caunterbury wolden ryde. The chambres and the stables weren wyde, And wel we weren esed atte beste. And shortly, whan the sonne was to reste, So hadde I spoken with hem everychon, That I was of hir felaweshipe anon, And made forward erly for to ryse, To take oure wey, ther as I yow devyse. But nathelees, whil I have tyme and space, Er that I ferther in this tale pace, Me thynketh it acordaunt to resoun To telle yow al the condicioun Of ech of hem, so as it semed me, And whiche they weren and of what degree, And eek in what array that they were inne; And at a Knyght than wol I first bigynne. A Knyght ther was, and that a worthy man, That fro the tyme that he first bigan To riden out, he loved chivalrie, Trouthe and honóur, fredom and curteisie. Ful worthy was he in his lordes werre, And thereto hadde he riden, no man ferre, As wel in cristendom as in hethenesse, And evere honóured for his worthynesse. At Alisaundre he was whan it was wonne; F…
November 28, 2024 at 12:28 AM
growth.wtf
Elon musk tweeting "Concerning"
November 28, 2024 at 12:08 AM
growth.wtf
He blocked me after this :-(
November 27, 2024 at 7:02 AM
growth.wtf
Explaining the watershed algorithm
November 23, 2024 at 8:23 PM