Are We Prompting New AI Models the Wrong Way?
The Daily AI Show
1 DAYS AGO
Are We Prompting New AI Models the Wrong Way?
Are We Prompting New AI Models the Wrong Way?
The Daily AI Show
1 DAYS AGO
The hosts return to discuss a range of AI topics, starting with a new security plugin for Claude Code that automates threat modeling. This sparks a broader conversation about local machine security, the differences between Mac and Windows, and the challenges of building custom security tools. The discussion then shifts to practical AI adoption, examining a Google report on how AI is used as an assistive tool across various industries, including manual trades like auto repair.
The episode explores Anthropic's Claude Code security plugin, which uses agents to map code bases and build threat models with independent review, leading to a discussion on local security tools like CCleaner and malware detection. The hosts then analyze Google's AI and Economy Atlas, which finds AI is used more for task assistance than full automation, extending beyond white-collar jobs into manual and technical fields like automotive repair. They discuss how AI can upskill tradespeople through multimodal diagnostics and smart glasses. The conversation concludes with a practical tip: newer reasoning models perform better with shorter, goal-oriented prompts, as demonstrated by Claude Code reducing its system prompt by 80%. The hosts also highlight the Language Model Builder tool for learning AI by building custom small models.
00:00
00:00
Anthropic released a security plugin for Claude Code
01:41
01:41
Ensure code doesn't introduce security vulnerabilities
02:03
02:03
Plugin analyzes code bases for vulnerabilities
03:25
03:25
Could Claude Code replace CCleaner?
05:06
05:06
Combine deep research with security plugins
06:02
06:02
Macs are not inherently safer than PCs
07:00
07:00
Self-awareness of what one can't replicate
09:49
09:49
Kimi K3 built via distillation of Fable 5
09:58
09:58
Something was accomplished in a particular manner.
10:00
10:00
An attempt to discredit Moonshot's achievement
11:17
11:17
The claim may be exaggerated, but there is evidence of a known pattern.
12:11
12:11
Traditional distillation from Fable 5 would have been too quick to be plausible
13:20
13:20
Perplexity no longer feels like AI slop
14:01
14:01
These policies are not new
15:16
15:16
Banning open models for citizens is legally and practically challenging
16:05
16:05
Local use is technically possible but resource-prohibitive
17:09
17:09
Canada often faces more company-level restrictions
18:40
18:40
Strix has 43,000 stars and connects to an AI API
19:10
19:10
Highly rated but expensive tool
19:37
19:37
AI code must be hardened against health data theft
20:43
20:43
Google's new AI and Economy Atlas report
22:03
22:03
Only a third of its data comes from English-speaking U.S. users.
23:00
23:00
AI serves as a live collaborator for diagnostics.
24:12
24:12
AI adoption is broad across industries and occupations
25:33
25:33
AI is more a collaborator than an automation tool.
27:04
27:04
Demand for radiologists has increased despite warnings
27:43
27:43
AI can aid rather than replace skilled workers
29:04
29:04
AI augments, not replaces, human expertise
30:06
30:06
AI will not replace mechanics but requires upskilling.
33:00
33:00
AI could replace the mechanic
33:40
33:40
Modern cars are heavily computerized, requiring caution.
34:04
34:04
AI-assisted video tutorials make complex tasks accessible
36:10
36:10
Proactively replace the 12-volt battery to avoid being stranded.
37:35
37:35
Claude Code reduced its system prompt by 80%.
37:36
37:36
Newer models prefer fewer examples to avoid constraining behavior.
38:00
38:00
Higher-end models are negatively impacted by overly verbose prompting
39:01
39:01
Shorter prompts improve AI output quality
40:05
40:05
Too many instructions create cognitive load.
41:04
41:04
Use fewer, more general instructions for newer models
42:42
42:42
Claude Code reduced its system prompt by 80%
43:31
43:31
Less specific instructions might have avoided the issue
44:04
44:04
Build systems for repeatable prompting, not just prompts.
45:07
45:07
Focus on building a structured prompting system
46:38
46:38
LanguageModelBuilder.com offers educational content on LLMs.
47:59
47:59
Start a project with various base models
48:15
48:15
Bring your own data to build your own model
48:38
48:38
Build your own language model.
49:35
49:35
Learn AI by building it yourself
50:33
50:33
Everyone will want to create their own custom models.
51:03
51:03
Train a GPT-2 small class model in a day
52:04
52:04
Customized data may not match past AI performance
53:01
53:01
Fine-tuning AI models on Macs is currently M-series only.
54:40
54:40
Gradient descent reduces loss to optimize AI training weights.
56:29
56:29
Using language models to clone oneself from emails
57:24
57:24
AI helps me be a better version of myself.
59:24
59:24
AI models need less verbose communication
1:00:02
1:00:02
Being teased means you're trusted and loved
1:01:35
1:01:35
Goodbye until next day
