fact check94% confident⏱ 1.7s
Llama 3 was trained on 15 trillion tokens, not 13.
Meta's announcement post specifies a 15T-token training corpus — roughly 7× Llama 2's. Source confirms the figure.
ai.meta.com
Ghost MicBetaTry now →A live fact-checker for podcasts, panels, and debates.
Ghost Mic listens while you talk and shows source-backed fact-checks on your screen in under two seconds. Optimized for speed using the Cerebras Inference API, it is blazing fast—delivering instant context without breaking your conversation's flow.
app.ghostmic.net
Ghost MicBetaLive transcript
So Meta dropped Llama 3 last spring — huge jump from Llama 2.
The context window doubled and the training run was massive.
Llama 3 was trained on about 13 trillion tokens.
Which is why the benchmark gains looked so dramatic…
Cards
Who it's for.
Podcasters
Stay accurate without breaking the flow.
Public speakers
Have your sources at hand while you present.
Debaters
Catch the slip the moment it happens.
Anyone in a real discussion
From dinner-table arguments to panel Q&As.
How it works.
- Step 1
Start talking
Open the app and begin your conversation. Ghost Mic listens.
- Step 2
Glance at the cards
Fact-checks, charts, perspectives, and sources appear on the side. Save the good ones. Ignore the rest.
What you get.
- Blazing-fast live fact-checks with sources in under two seconds, powered by Cerebras Inference
- Save what's useful, dismiss what isn't
- Works in your browser, nothing to install
- Free for a limited time during beta
FAQ.
Is my audio recorded?+
No. Transcripts are processed in memory and not stored. Raw audio never touches disk.
Will it interrupt me?+
No. Ghost Mic only shows cards on screen — it never speaks.
What languages?+
English for now.
What if it gets a fact wrong?+
Every card cites a source you can verify in a click. Confidence is shown on each card, and we skip the call when the model can't ground it.
Does it work with Zoom, Google Meet, or Riverside?+
It listens to your mic. Use it alongside any recording tool.
Is it really free?+
Yes — Ghost Mic is free for a limited time while we're in beta. No credit card required. We'll give plenty of notice before anything changes.
Screenshot gallery — remove before launch
Capture each mock below. Crop out titles if needed. Revert this section when done.
Catch factual slips before they become clips.
Ghost Mic listens for checkable claims and surfaces a source-backed correction while the conversation is still live.
app.ghostmic.net
Ghost MicBeta● recordingLive transcript
So Meta dropped Llama 3 last spring — huge jump from Llama 2.
The context window doubled and the training run was massive.
Llama 3 was trained on about 13 trillion tokens.
Which is why the benchmark gains looked so dramatic…
Cards
fact check94% confident⏱ 1.7s
Llama 3 was trained on 15 trillion tokens, not 13.
Meta's announcement post specifies a 15T-token training corpus — roughly 7× Llama 2's. Source confirms the figure.
ai.meta.com
Correct numbers in real time.
When a stat is off, Ghost Mic gives you the sourced number so you can correct yourself on air.
app.ghostmic.net
Ghost MicBeta● recordingLive transcript
Let's talk about the latest jobs report — mixed signals overall.
Unemployment ticked up slightly but hiring stayed resilient.
The US added about 350,000 jobs in March alone.
That's way above what most economists were expecting…
Cards
fact check91% confident⏱ 1.9s
BLS initially reported 303,000 March jobs, not 350,000.
BLS initially reported March 2024 nonfarm payrolls rose by 303K; the figure was later revised to 310K.
bls.gov
Add instant context for names your audience may not know.
Drop a company, paper, product, or person into the conversation and get a quick primer with a source.
app.ghostmic.net
Ghost MicBeta● recordingLive transcript
Inference speed is the new moat for AI apps.
Groq made a splash with their LPU, but Cerebras is pushing hard too.
They claim wafer-scale chips beat GPU clusters on tokens per watt.
We should dig into whether that's lab-only or production-ready…
Cards
context88% confident⏱ 1.5s
Cerebras builds dinner-plate-sized AI chips.
Cerebras Systems makes the Wafer-Scale Engine — a single chip the size of a wafer, designed for training and inference at datacenter scale.
cerebras.ai
Turn jargon into plain English.
Acronyms and technical terms get translated into a short, readable definition your audience can follow.
app.ghostmic.net
Ghost MicBeta● recordingLive transcript
Alignment is the whole game now — safety, helpfulness, tone.
Every frontier lab is iterating on RLHF and its successors.
Constitutional AI and DPO get name-checked a lot too.
But listeners might not know what any of that actually means…
Cards
definition92% confident⏱ 1.4s
RLHF = humans rank model outputs to train preferences.
Reinforcement Learning from Human Feedback: annotators compare answers, and the model learns to prefer higher-rated responses.
arxiv.org
Find the next angle when the conversation stalls.
When there is a lull, Ghost Mic can suggest a timely follow-up backed by something concrete.
app.ghostmic.net
Ghost MicBeta● recordingLive transcript
EV sales are up again quarter over quarter, no surprise there.
The charging network is still the pain point in rural areas.
Yeah… [pause] …not sure where to take this next.
Maybe we pivot to grid capacity? Or battery recycling?
Cards
topic76% confident⏱ 1.6s
Ask how EV charger tax credits are changing.
The federal 30C credit can cover EV charging equipment in eligible locations, but phaseout dates and eligibility vary for homes and businesses.
irs.gov
Works cleanly on mobile, too.
Cards stay on top while the live transcript sits in a compact drawer at the bottom.
app.ghostmic.net
Ghost MicBeta● recordingCards
fact check89% confident⏱ 2.1s
Altman's OpenAI ouster lasted less than five days.
OpenAI announced Altman's removal on Nov. 17, 2023 and reached an agreement for his return late Nov. 21 / Nov. 22, depending on timezone.
time.com
Live transcript
Quick take on the OpenAI board saga…
Sam was out for nearly a week before the reinstatement.
Most employees threatened to quit if he didn't come back.