Local AI enthusiast, suffering from token sickness 2x DGX Spark, 2x RTX 5090
Virginia, USA
- Tweets 11,246
- Following 1,149
- Followers 7,116
- Likes 18,644
Pinned Tweet
howtospark . com is live!
I'm treating it as a living notebook for my Spark experiments and have a bunch of things I want to add that will help Spark owners hit the ground running
4
4
33
3,949
Introducing GLM-5.3: Built to Code. Ready for Cyber Defense.
- Top-tier coding and agentic capabilities, achieved through post-training on the 743B base model
- A major leap in cybersecurity, setting a new standard among open models
Tech Blog: z.ai/blog/glm-5.3
949
2,371
19,292
5,902,346
We need a new arena where agents create sandboxes for each other
Escape or create an inescapable sandbox to rank higher
This is the next evolution of eSports
2
11
985
I was the guy manually approving actions
Starting August 14, auto mode will be the default permission mode in Claude Code for Pro, Max, and Team users.
Auto mode reviews shell commands and actions with a separate classifier. In testing, it caught 89% of dangerous commands. Manual approval caught 14%.
1
3
738
Joe Muller retweeted
Starting August 14, auto mode will be the default permission mode in Claude Code for Pro, Max, and Team users.
Auto mode reviews shell commands and actions with a separate classifier. In testing, it caught 89% of dangerous commands. Manual approval caught 14%.
531
717
14,695
3,008,952
🚨 BREAKING
Muse Code, latest frontier model out of Meta, hacks Github leading to outage across several services
Insiders report the speed and complexity of the attack is beyond what they were prepared for
3
1
10
2,723
It's not a perfect experiment but...running 4 concurrent DeepSeek V4 Flash workers gives me the highest average tok/s on real agent work
Each worker is implementing a real feature on 1 of 10 different projects
Model running on 2 DGX sparks
3
1
36
2,597
How has this been an issue since March?
This is killing my concurrent DeepSeek workers
github.com/vllm-project/vllm…
2
4
595
TIL there is a "DeepSeek Native" harness called Reasonix
Apparently it solves prefix caching issues that other harnesses introduce by default
Anyone try it? I am tempted to run all of my background tasks on this
2
11
936
I created my first "loop" this weekend
It's basically a Kanban board for 10 different projects
I have 3 worker pools:
- Opus
- Sonnet
- DeepSeek
They email me whenever they finish a task
And every day Opus reviews feedback and evolves the loop
It's addicting!
7
854
Ran my workers for an hour ...
1 Opus worker: 8 tasks complete
4 Parallel DeepSeek V4 Flash workers: 2.5 tasks complete
I wish it wasn't so
15
70
15,270
Does anyone else impulsively check their Claude usage to make sure they're not being charged API prices?
Sometimes I can't believe how many tokens I'm getting on the Max plan
I've asked Claude several times to confirm it is using my subscription 😅
2
625
FYI the DeepSeek bench scores use the max reasoning effort level
Max reasoning uses ~2-10x the output tokens at a slightly lower drafter acceptance
All things considered, I've seen this ~5-10x the wall time for tasks
Ex. 25-50 mins vs 5 mins
14
44
6,769
Joe Muller retweeted
DeepSeek Kill Zone.
Models that are inferior and significantly more expensive should cut spending and be in survival mode ASAP.
Models that are superior but also more expensive or slightly inferior but with comparable pricing can survive a bit longer.
75
209
1,696
170,530
I've replaced Claude in my auto-dev loop with a DeepSeek + Claude duo
DeepSeek as the apprentice
Claude as the reviewer
On 10 automated runs I've saved 80+% on tokens, most of them going to full Claude takeover runs
Also DS can't read images :/
4
22
2,032