Aaron π οΈπ€π retweeted
The removal of rollover credits in AI should be treated as theft @FTC
Millions of dollars per month are effectively being taken from paying customers simply because they didnβt use the credits in time.
Over 100,000 credits were removed from my @ManusAI account after I spent thousands of dollars acquiring them.
This is fundamentally different from a flat-fee structure like Claude.
Which is closer to a cable subscription where Iβm paying for access to the platform, not for a specific quantity of usage. If I donβt watch any shows, I donβt get a refund.
Manus is different. It is an AI wrapper, not the underlying LLM provider.
When I buy credits on a wrapper, those credits represent a pre-ordered quantity of tokens/usage. If I pre-order tokens directly through the Anthropic API and never use them, the balance remains.
I regularly buy credit packs in the Anthropic console; the dollar amount only decreases when actual token usage occurs.
As a wrapper, Manus incurred no cost for the 100,000+ credits that were later removed, because no usage ever happened. Yet, I still paid the full amount.
When a company sells me X credits that map to Y amount of usage... and it never had to pay for any of that usage because it is only a wrapper... those credits should remain mine until they are actually converted into service.
You incurred no (or very minimal) cost, so I should not lose all of the value I paid.
Taking the money while delivering none of the prepaid product or service is not clever billing. It is theft.
Iβve read the terms. You explicitly treat add-on credits differently, yet the ability to purchase true add-on credits has now disappeared. π€
The only option is to upgrade to a higher monthly plan, which immediately puts those same credits at risk under your own policy.
Just because the wording in the contract can be framed as βstandardβ does not make it legitimate.
The fact is, I paid for a defined amount of usage, no usage (and no corresponding cost) was ever incurred by the company, but the credits were still taken.
Since AI wrappers are still quite a new business model I would really value the perspective of people like @garrytan @jasonlk @BostonVC
Am I egregiously wrong here and owe Manus a public apology, or is my logic sound and something is wrong here. This is not the first AI wrapper I've encountered a loss like this from, just the biggest. 100k credits is a substantial amount of work that could be done.
Why Iβm actually recommending the DGX Spark now, after advising against it 3 months ago:
Memory bandwidth was always the main bottleneck for local setup choices. Speculative decoding changed that completely.
On paper, a DGX Spark at 273 GB/s running DeepSeek V4 Flash (13B Active) tops out at just 21 tok/s.
Speculative decoding bypasses this wall entirely. You get speeds multiple times faster, actually beating top frontier models.
Early concerns about output quality loss are gone as well. It runs with almost zero degradation now.
Local AI Era is here.
Wowβ¦ Iβm actually blown away by how far AI video editing has come.
If you ever wanted to start a YT channel, now is the time.
I recorded this video and Codex one shot the edit with Hyperframes and GPT 2.
A friend and I are building this app and channel for fun in our spare time.
This is what my process wasβ¦
1. Create a brand first. Have GPT image 2 create a logo, then once you like it, have it create a full brand kit in PDF and Markdown Format
2. Download PromptSmart Pro app on your iPhone. Itβs $30, but worth it.
3. Pick a topic and have GPT write a script thatβs not beyond a 7th grade reading level. (This keeps it from going off the rails with AI speak)
4. Use PromptSmart to record yourself reading the script.
5. Send GPT the finished video and tell it to install Hyperframes and use the Hyperframes skills to edit the video according to your script and use the brand kit to keep the visuals consistent. Also tell it to use GPT image 2 to make slides that would be helpful to your video.
6. When the video is complete, tell it to create a thumbnail, Title, description, and tags, then upload the video to YT.
You can even have it splice the video into shorts for YT and other platforms.
This is a great way to make decent (not perfect) videos that are AI assisted but still authentic and not AI generated junk.
Itβs pretty fun too.
youtu.be/OpaNBDhBo1E?is=jHdnβ¦






