Are We Using Opus 5 Wrong?
The Daily AI Show
1 DAYS AGO
Are We Using Opus 5 Wrong?
Are We Using Opus 5 Wrong?
The Daily AI Show
1 DAYS AGO
This episode of the Daily AI Show covers the mixed early reactions to Opus 5, the debate over open-weight AI models, and practical tips for getting better results from newer reasoning models. The hosts also explore the new ChatGPT Sites beta, demonstrating its use for building dashboards, internal tools, and sales prototypes, and discuss the potential of ChatGPT Voice for richer, more natural interactions.
The hosts discuss Opus 5's low cost and high performance, but note that old, overly detailed prompts may limit its effectiveness. They advise stripping back restrictive instructions to allow the model to explore 'known unknowns' for better results. The conversation shifts to the Open Weight AI Alliance, with Dario Amodei clarifying Anthropic's stance against a blanket ban, instead advocating for targeted restrictions on chip sales and model distillation to authoritarian states. The second half focuses on ChatGPT Sites, where the hosts demonstrate building a searchable podcast website and a business intelligence dashboard. They highlight the ease of one-click publishing but warn about privacy risks and platform lock-in, recommending GitHub for portability. Finally, they explore ChatGPT Voice, noting that spoken prompts add nuance and context, making it a powerful tool for dictation and a 'Jarvis-style' orchestrator, despite using more tokens than typing.
00:00
00:00
Opus 5 is praised as the best current model
02:35
02:35
Half the price of Opus and a significant leap forward
04:30
04:30
Restrictions would harm US open models
06:19
06:19
Authoritarian governments could use AI for military superiority or repression
07:31
07:31
Risks of powerful models being used for cyber or biological attacks
08:11
08:11
Crack down on industrial-scale model distillation
08:34
08:34
Anthropic does not support a blanket ban on open-weight models
09:13
09:13
Pre-release safety testing is key for cyber, biological, and alignment risks.
12:37
12:37
Closing loopholes would help open models
17:09
17:09
Old, overly detailed prompts may not work well with Opus 5
19:27
19:27
Skills are just markdown files with instructions.
19:42
19:42
Adapt prompting skills to unlock Opus 5's potential
20:35
20:35
Stripping old skills reduced hallucinations
22:07
22:07
Removed something to improve accuracy and clarity
22:46
22:46
Building a better version of Salesforce reports
23:35
23:35
ChatGPT Sites builds personalized BI dashboards.
27:28
27:28
The actual episode count is 777, not 849.
28:23
28:23
Search pulls from YouTube data without API validation
30:02
30:02
Using Claude to generate a dedicated episode page
32:19
32:19
Two paths: one-click publish or 30-minute move to Netlify
34:02
34:02
Free hosting locks users in without GitHub backup
35:11
35:11
Sensitive content can be unintentionally exposed.
36:20
36:20
OpenAI may restrict public access if usage becomes unusually high.
38:08
38:08
Most sites get few visits.
39:53
39:53
Anthropic fixed the no-index tag and removed exposed chats
41:50
41:50
Build a demo during a call using dummy data.
42:54
42:54
ChatGPT Sites boost sales close rates.
45:06
45:06
Desire for a Jarvis-style AI orchestrator
45:15
45:15
ChatGPT Voice acts as a Jarvis-like orchestrator.
47:31
47:31
A boom mic would encourage more frequent voice use
48:28
48:28
Silent background listening is a key convenience.
50:44
50:44
Voice dictation is faster and less mentally taxing than typing
51:30
51:30
Advanced models pick up spoken nuances lost in written prompts
52:02
52:02
Voice notes capture more nuance than typed drafts.
54:49
54:49
I prefer British voices for AI, even though my mother is Australian.
59:25
59:25
Functionality matters more than aesthetics
1:01:17
1:01:17
If it works and solves a problem, the process doesn't matter.
1:04:59
1:04:59
Thank you for your support.
