iiinova — where your digital self keeps evolving

Philip Nuzhnyi @callmephilip.com · 07/10/2026

"This paper was written with human assistance." - github.com/openai/math/...

Philip Nuzhnyi @callmephilip.com · 18/09/2026

Get rid of any results that are not our supported choice indices. Turn logits (random looking numbers like 3, 89, -386) into a list of probabilities that add up to 1. Get your answers with probabilities back.

000

Philip Nuzhnyi @callmephilip.com · 18/09/2026

! Do 1 round through the model (as opposed to autoregressive loop where you feed results from the previous batch into the next iteration and keep assembling final result). Only grab logits for the last chunk (what comes after "choice_index:" in our case). Important: send multiple questions in 1 go

100

Philip Nuzhnyi @callmephilip.com · 18/09/2026

Massage your prompt so it looks like "user says / assistant says" with our prefix at the very end so that the model is set up to spit out a single value for one of the choices.

100

Philip Nuzhnyi @callmephilip.com · 18/09/2026

Figure out model's internal encoding representation for the IDs your system supports (up to 255 choices in this case). This extra PREFIX business seems to be dealing with possible different ways to encode the same value (e.g. "2") depending on whether it appears in "2" or "choice_index: 2"

100

Philip Nuzhnyi @callmephilip.com · 18/09/2026

Each question is framed as a _numbered_ multiple choice question. E.g. Options like "blue, red, yellow" become 0, 1, 2. The model will answer with the number.

100

Philip Nuzhnyi @callmephilip.com · 15/09/2026

Seems to be working directly from mobile inside an RN app after applying some elbow grease - pdsls.dev/at://did:plc...

010

Philip Nuzhnyi @callmephilip.com · 17/07/2026

"Momentum is the supreme law" - Antoine Bourdelle // By Antoine Bourdelle - Self-photographed / PierreSelim, CC0, commons.wikimedia.org/w/index.php?...

000

Philip Nuzhnyi @callmephilip.com · 16/07/2026

- Generator produces reasoning trajectories for new queries (strategies and recurring pitfalls) - Reflector critiques these traces - Curator synthesizes these lessons into compact delta entries - deltas get merged deterministically into the existing context

The framework of Agentic Context Engineering (ACE).

100

Philip Nuzhnyi @callmephilip.com · 09/07/2026

Been seeing a lot of this lately. None of my jazz is cybersecurity adjacent. Their classifiers need to improve

000

Philip Nuzhnyi @callmephilip.com · 30/04/2026

Peak Pickle, LFG - reidhoffman.substack.com/p/in-defense... via @reidhoffman.bsky.social img from godfatherofmath.blogspot.com/2013/05/befo...

image of heinz pickle sign from early 1900s

010

Philip Nuzhnyi @callmephilip.com · 14/04/2026

Filing this under universal basic income tag

010

Philip Nuzhnyi @callmephilip.com · 14/04/2026

Apr 2026 check-in on the local models via www.latent.space/p/ainews-top...

- Qwen 3.5 — most broadly recommended family right now across usecases.
- Gemma 4 — strong recent buzz for local usability, especially smaller and mid-sized deployments.
- GLM-5 / GLM-4.7 — near the top of broad open-model rankings, increasingly part of the “best overall” conversation.
- MiniMax M2.5 / M2.7 — repeatedly cited for agentic/tool-heavy workloads.
- DeepSeek V3.2 — still firmly in the top cluster when people talk about strongest open-weight general models.
- GPT-oss 20B — not the mainstream “winner,” but increasingly recommended as a practical local option and for uncensored variants.

240

Philip Nuzhnyi @callmephilip.com · 13/02/2026

Zoo - zoo.dev "a conversational CAD agent for reasoning-driven design on a modern geometry engine."

100

Philip Nuzhnyi @callmephilip.com · 01/02/2026

"One major aspect of this extended mind is our capacity for interoception - an awareness of the internal sensations of the body that gives us a kind of non-conscious knowledge, which we might term intuition but which is rooted in concrete embodied experience and somatic pattern-recognition."

000

Philip Nuzhnyi @callmephilip.com · 25/01/2026

21/* "The saucy way of putting it is that human beings [...] are apes with access to linguistic recursion [...] plus cultural evolution that has allowed us to amass an ever-growing pile of complex sticks that we can bang together"

010

Philip Nuzhnyi @callmephilip.com · 14/01/2026

Interesting map from the Economist. A great reminder how powerful framing effect is

020

Philip Nuzhnyi @callmephilip.com · 07/11/2025

Apparently abiogenesis is also sometimes referred to as "biopoesis" - which is a. poetic b. poesque

020

Philip Nuzhnyi @callmephilip.com · 03/11/2025

"Each program inside tmux gets its own terminal managed by tmux, which can be accessed from the single terminal where tmux is running - this called multiplexing and tmux is a terminal multiplexer." - from github.com/tmux/tmux/wi...

010

Philip Nuzhnyi @callmephilip.com · 11/10/2025

My buddy Dostoevsky is cycling from Lisbon to Malaga and I am rawdogging it! whereisdostoevsky.com source: github.com/callmephilip...

000

Philip Nuzhnyi @callmephilip.com · 22/09/2025

Frenchies took over MS's docs department - www.typescriptlang.org/docs/handboo...

000

Philip Nuzhnyi @callmephilip.com · 09/09/2025

Claude Code layered interactions (source: blog.promptlayer.com/claude-code-...)

100

Philip Nuzhnyi @callmephilip.com · 09/09/2025

Claude Code Flowchart (source: blog.promptlayer.com/claude-code-...)

100

Philip Nuzhnyi @callmephilip.com · 11/08/2025

It's week 3. Claude is not helping at all

010

Philip Nuzhnyi @callmephilip.com · 08/08/2025

alias claudius="claude --dangerously-skip-permissions"

010

Philip Nuzhnyi @callmephilip.com · 08/08/2025

Interesting typographical choices here (from en.wikipedia.org/wiki/COBOL)

100

Philip Nuzhnyi @callmephilip.com · 04/08/2025

It's not enough to just throw a bunch of tools at the model and hope for it to "get it". LLMs need to be explained when and how to use Claude's system prompt breakdown from www.dbreunig.com/2025/05/07/c...

010

Philip Nuzhnyi @callmephilip.com · 03/08/2025

TIL: all those whimsical gerunds in Claude code get generated on the fly via Haiku. I am lost for words. Should have asked Haiku for an appropriate verbal concoction to use

010

Philip Nuzhnyi @callmephilip.com · 25/07/2025

You can subclass envs to create stricter versions of them, for example

000

Philip Nuzhnyi @callmephilip.com · 25/07/2025

You can hijack execution for specific actions

000

Philip Nuzhnyi @callmephilip.com · 25/07/2025

Here's the LOOP: - call LLM using current messages - parse the response and get the action - exec the action - render observation (`DefaultAgent.render_template`) - add observation to the message list

100

Philip Nuzhnyi @callmephilip.com · 25/07/2025

- every action is executed via `subprocess.run`: each action is independent from the previous (there is no central shell session). Why is this a big deal? ⤵️

000

Philip Nuzhnyi @callmephilip.com · 23/07/2025

They: your product is just a Chat GPT wrapper You: genius.com/Chase-and-st...

010

Philip Nuzhnyi @callmephilip.com · 19/07/2025

Trying to get Claude to give me a nice ASCII art style Winnie-the-Pooh. I mentioned Claude Monet (half jokingly) and things escalated.

020

Philip Nuzhnyi @callmephilip.com · 17/07/2025

Marcel chauffe! 😂 love it. `marcel grève` is a chef's kiss github.com/brouberol/ma...

100

Philip Nuzhnyi @callmephilip.com · 05/07/2025

Watched “High noon” last night. Here is the old marshal explaining Cooper why no one is willing to help him defend their town: “People gotta talk themselves into law and order before they do anything about it. Maybe because down deep they don't care. They just don't care.”

000

Philip Nuzhnyi @callmephilip.com · 04/07/2025

"Context Offloading is the act of storing information outside the LLM’s context, usually via a tool that stores and manages the data." Create a "scratchpad" - a place for the model to write down notes keeping them out of the context. Offloading scenarios (from Anthropic):

000

Philip Nuzhnyi @callmephilip.com · 04/07/2025

"Context Pruning is the act of removing irrelevant or otherwise unneeded information from the context." Example using Provence:

100

Philip Nuzhnyi @callmephilip.com · 04/07/2025

"Context Quarantine is the act of isolating contexts in their own dedicated threads, each used separately by one or more LLMs." Break up tasks into smaller, isolated jobs with their own contexts re: www.anthropic.com/engineering/...

100

Philip Nuzhnyi @callmephilip.com · 04/07/2025

"Tool Loadout is the act of selecting only relevant tool definitions to add to your context." The most basic approach - vectorize tool descriptions and then search for tools.

100

Philip Nuzhnyi @callmephilip.com · 04/07/2025

It took me a while to finally get my hands on uv. Sounds like there's no more pip for me - github.com/astral-sh/uv

000

Philip Nuzhnyi @callmephilip.com · 04/07/2025

"Context Distraction is when a context grows so long that the model over-focuses on the context, neglecting what it learned during training." Agent repeats past actions from its huge ctx history. This is the case with big models, even more so with smaller ones (www.databricks.com/blog/long-co...)

100

Philip Nuzhnyi @callmephilip.com · 04/07/2025

Did traditional programming condition humans to act more like machines and aren't LLMs liberating us (developers) from this conditioning then? Instead of "coding" as identity, let's do "building stuff people actually want". from corecursive.com/coding-agent... h/t @simonwillison.net

000