developer/programmer/ai nerd/🍁
Canada
Joined October 2023
- Tweets6.3K
- Following2.4K
- Followers550
- Likes10.6K
Pinned Tweet
So I was bored tonite and thought I'd make a minecraft meets space travel type of game. Its kind of how I envision Minecraft 2 should be. Anyway I made it in threejs and might work on it more as it is actually not that bad of a start. What do you think random internet people?
Pretty easy to find the shill accounts to block. Usually start with "Holy crap".. or "Holy Sh***" and then go gaw gaw over some crap product
39
All this lame drama by the AI companies is probably just a publicity stunt in disguise. Gotta keep the camera's focused to ensure those investment bucks stay interested...
1
41
We are now in the age of information. Let us not mourn the jobs that are lost through obsolescence, but focus on celebrating the new jobs/opportunities that are yet to come, that will propel us into the future.
2
39
Gatekeepers gatekeeping is not going to help society progress. True science embraces change.
Mathematicians are coping so hard. Suddenly they can not stand AI. 😀
the guild is so badly protecting its turf.
The Association for Human Mathematics (AHM), criticised OpenAI's yesterday's release of 722 math manuscripts and urged mathematicians "to discontinue their work with OpenAI".
And then Terence Tao's reposted AHM's statement on his blog.
Walking away is not a scientific position.
2
58
An 80 yr old lady alone in her house, is she better protected by a government promise to protect her, or is she better off with a 12 gauge shotgun against 3 linebackers breaking into her home? Maybe neutering the models is not protecting anyone properly, and only making victims?
1
32
It should be called 'Augmented Intelligence' because by itself, it is useless, but commanded, it boosts the capabilities of the individual commanding it.
33
Nothing quite sells chinese AI quite like the behaviour of OpenAI and Anthropic towards their customers.
3
50
g023 Code 1.3.4 is out.
A pure-Python coding agent for your terminal, powered by DeepSeek V4. Subagent-first. Context is currency. One model, one endpoint: deepseek-flash on POST /responses.
The last cut was about the price sheet. This one is about the machine that spends it: an orchestrator that hires specialists, lets them talk both ways, retunes how hard it thinks, and — when the mission is actually done — folds the transcript into a one-page brief for the next job.
THE ORCHESTRATOR IS A LOOP, NOT A CHAT WRAPPER
One user prompt becomes as many model→tools cycles as needed. The orchestrator never dumps a whole file into its own history unless it asked for a line range or raw=true. Heavy I/O happens in isolated workers: ReadFile, SearchContent, AnalyzeImage, Agent. Compact JSON comes back. Parent context stays a briefing, not a warehouse.
Think of it as a miser with a gold card: it will spend, but it would rather send a specialist than haul the whole project onto the table.
FOUR SPECIALISTS, DEPTH ONE
The Agent tool hires four kinds of worker:
• research — inspect the tree, relay findings
• planning — inspect, optionally from prior research, relay a plan
• todos — write a list, then orchestrate the task workers (task workers fix prior worker's issues surfaced and then complete their objective)
• task — implement ONE todo (this kind can edit, write, and run a shell)
Research and planning are read-only. Task is the one that touches the tree. LearnPut and Agent are refused at every kind, so a worker cannot spawn a grandchild. Depth-2 was measured and not shipped: a second hop is another miss on a new input prefix plus another hop of output.
THEY COMMUNICATE BOTH WAYS — IN JSON, NOT IN A NESTED CHAT
The schema says it in those words: pass prior/issues so workers communicate both ways; the parent never ingests a nested transcript.
Parent → worker: the objective, plus any prior payload, plus issues from the last task.
Worker → parent: a size-capped JSON object (findings, plan, result, issues). Not the nested homework. Cap is 8 model calls and 12,000 characters.
Worker → next worker: research feeds planning via prior; planning feeds todos via prior; in the todos pipeline, task N's issues become task N+1's brief.
The nested loop is isolated. The parent does not talk to the child mid-loop. "Both ways" is the relay at spawn and return, plus the sequential issues handoff between task workers. Each run also writes a branch journal under .g023/scratch/agent/ — kind, objective, payload, per-item outline, not the full transcript — so offload is inspectable after the fact.
Nested workers used to miss from token 0 on a different system prompt and a subset tool list. Now they reuse the parent's frozen prefix (instructions + tools + pinned effort). The worker brief moves into the first user item. Execution still refuses any tool the kind does not allow. Standing cache hits; isolated input still misses. That is the fork.
DYNAMIC THINKING
Thinking is a 0–100 dial, default and ceiling 75 (wire alias high). max / 100 / xhigh cap to 75: on this endpoint they often burn the entire output budget on thinking and return an empty answer. Reasoning tokens are billed as output (~200× a cache hit). That is why 75 is the ceiling, not a taste preference.
On the same final answer — after the user-visible text, never shown to you — the model may write two trailers:
[handoff]
## Goals / Files / Decisions / Done
[/handoff]
[next_effort: N]
N is how hard the NEXT user request looks, not this turn's remaining tool steps. 25–50 is routine follow-up; 0 disables thinking. Pick the number that matches the next request; do not default to 75.
The dial is not retuned mid-turn. Changing reasoning.effort on a hot prefix misses from token 0. The stored number applies when the prefix is already new: /thinking, /goal, /clear, or a successful post-turn compact. /goal forces 75 for its own turn and keeps a same-response [next_effort] for the next request instead of restoring the pre-goal dial.
WHEN THE MISSION IS DONE, THE TRANSCRIPT DOES NOT RIDE ALONG
After a finished turn, post-turn compact may replace history so the next request is standing instructions + working-set item + a one-page brief — not the tool-call novel.
If the working-set judge says the objective is complete, that compact always fires, even on a hot prefix. The next request is a new orchestration. Summary source, in order: the [handoff] already on this turn's final answer (no second Flash call); a local fold of an existing compact-summary item; otherwise Flash. On success it writes .g023/logs/{stamp}.log for --continue / /resume log, and releases the pinned effort so a stored [next_effort] can retune on the new prefix.
"Done" is not the model's last sentence. A local judge is the only completed / not-completed verdict: no open leaves, and every done leaf cites a file hash or a passing test. Pending, empty, or done-without-evidence: not completed. Remaining work lives in the working-set object. Compact cannot summarise the todos away.
Worth naming: compact also fires when the window is full, or leftover bulk is large and uncached. A hot prefix with an open objective is cheaper to keep than to shrink. Layer 1 blanks aged tool dumps in place, only under window pressure. Layer 3 summarises the middle while a job is still open, keeping the sticky goal and a pairing-intact tail. Those are different levers. Mission-complete is the handoff into the next job.
WHY THIS SHAPE
Off-peak Flash (DeepSeek, 2026-09-10): cache hit $0.003/M, miss $0.15/M (~50×), output $0.60/M (~200× a hit). Peak is 2×, Monday–Friday only. The orchestration is built around those ratios: isolate the heavy work, relay JSON, pin effort on a hot prefix, compact only when the job is done or the leftover is uncached. End-to-end session cost is still not A/B'd. We say so. /cost is the meter. 421 tests, all offline, all passing.
STILL HERE, STILL THE ONES YOU FEEL
/surge converts Beijing peak windows to your clock. /or is OpenRouter vision. /orsub offloads Agent turns. /learn and .g023/memory are searched on demand, never prepended to the system prompt. Edit fails closed. WriteFile and Edit snapshot .g023/backups/ first; /undo restores. File tools refuse paths outside the project root.
Python 3.11+. One DeepSeek key. Run installer.sh (Windows: installer.bat). Type g023 from any project.
MIT. Subagent-first. Context is currency.
g023 Code 1.3.4 — github.com/g023/g023code
1
2
328
I can't abandon g023_code harness v1 as its just to awesome and different than v2, so I made some major improvements to v1 and plan on pushing a 1.3.3 release today here in the next little bit that incorporates some major token savings, and an even better end result.
75