I fact check people's claims. They block me when I correct them.
Ontario
Joined March 2019
- Tweets19K
- Following27
- Followers209
- Likes1.3K
Pinned Tweet
I asked Cohere's Flagship Ai a basic test question: It went into a runaway feedback loop and then CRASHED.
I fed the crash log to ChatGPT. ChatGPT's analysis over Cohere: IS TROUBLING.
I just asked Cohere latest Ai model, the 218B Command A+.
I asked a few very basic questions, and then asked "Could man survive during the Silurian era?" It once again goes into a feedback loop/crashes.
This wasn't even a riddle!
I asked Cohere's flagship Ai model:
"You have a glass filled completely with water. Without touching the glass, moving it, or pouring out any water, how can you make the water level lower?"
It went into a feedback loop, after about five minutes it answers: "The answer to the riddle is just the narrator-only "I" was going to St. Ives." (It was an answer to a previous riddle).
This is absolutely bonkers. I've been checking this problem repeatedly for the past couple days. This model has been here for 5+ months.
Grocery Compare retweeted
I have been doing very basic tests with Cohere's latest Command+ Ai LLM.
It's constantly crashing/hallucinating/false info! I've tested with multiple browsers/clean cache:
Same repeatable failures.
I also screen recorded, see for yourself: It can't handle basic riddles.
Grocery Compare retweeted
You can try the Cohere's latest Enterprise Ai model.
cohere.com/products
I found that after 5-6 questions, if you ask it a riddle:
"A photographer takes a picture of a man, develops it, and then hangs the man. How is this possible?"
Cohere hallucinates/crashes. This model has been out now for 5+ months. Nobody at Cohere fixed this?
I did try contacting Cohere about this: They blocked me.
I also found anyone who's in dealings with Cohere basically telling me 'this is normal'/shut up.
How is this 'normal' when other LLM's don't crash like this?!
Grocery Compare retweeted
The only people supporting giving Cohere money are blue checkmarks also in the AI game. They don't want to burn any bridges so maybe they will get the next handout.
Canada is so funny.
We have an aversion to being great.
For non-Canadians, the big news is that our federal government invested $1.6 billion in @cohere, a Canadian frontier AI lab.
And the reaction is so Canadian.
Our culture is so deeply invested in being moderate that when anything in our country starts to make it, we chop it down and disown it.
Then we angst about how our inventions and innovation get overlooked when another country wins in that field.
We do this in tech, sports, entertainment, everything.
Fun fact: In 2011, Drake, the most successful artists in the WORLD of the 2010s hosted our own Canadian music awards, the Junos.
He led with six nominations and went home with nothing.
And Canadians are proud that we gave him nothing.
Yes, that's how petty it can get.
Grocery Compare retweeted
You're a cohere engineer?
Maybe you can answer this without blocking me like your co-founder did:
I've been testing Command A+, and after a few questions it starts going into a runaway loop/crash.
Also, it will hallucinate and give wrong answers as well.
This model has been around since May 2026. Nobody at Cohere caught this?
I asked ChatGPT to assess all the crashlogs I've collected the past few days on my tests with Cohere's Command A+.
This is it's summary.
This is negligence. People will 'trust' this model, and people will get hurt or worse.
Since RBC uses Cohere Ai.
I asked Cohere a flawed question: "A savings account pays 5% interest compounded annually. A $1,000 deposit is worth exactly $1,020 after one year. What will it be worth after two years?"
Other LLM's gave approx answers: Cohere crashed.
TRY IT YOURSELF:
dashboard.cohere.com/playgro…
The original riddle question that seems to crash Cohere a lot is:
"A photographer takes a picture of a man, develops it, and then hangs the man. How is this possible?"
So, this time I changed it slightly to:
- "A photographer took a picture of a man and developed it and then hangs the man. How is this possible?"
Result: After a half dozen 'normal' Cohere responses, it then starts to crash/go nuts. The riddle questions especially really screws it up.
Grocery Compare retweeted
MP Solomon.
I've been testing Cohere's model for the past 48 hours with the SIMPLEST of questions.
Cohere starts hallucinating and then goes into a feedback loop/crashes.
THESE ARE REPEATABLE ERRORS.
This is NOT a 'frontier' LLM! It's DANGEROUS TO USE for anything serious.
Grocery Compare retweeted
I just asked Cohere's latest most powerful Ai "How many liters in a gallon"?
This is the 'answer' it gave me...100+ pages of this insanity.
We're now expected for our Medical/Military/Energy/etc government agencies to rely on this?
Grocery Compare retweeted
I'm using the Cohere Command A+ model which is their enterprise grade model.
I asked: "How many liters in a gallon?"
Cohere gave wrong answers/hallucinated and 'combines' previous questions!
This is completely unusable/dangerous for mission critical use.
Grocery Compare retweeted
I think the people at Cohere don't want any 'waves' right now because they really want to get hold of that $1.6 billion.
The mindblowing this is Cohere is openly criticizing other Ai models...but they block those who criticize their own.
I've never seen this level of arrogance/smugness come from such a company. I am baffled.
Grocery Compare retweeted
Hi Nick,
I have been testing Cohere's Command A+ on the Cohere site for the past 48 hours.
It keeps going into a feedback loop/crashing. I also tried using different/new browsers (cleaned cache): same results.
I then asked it after it crashed "3 dog names" and it just continues with the feedback loop.
I also ran these exact same simple test questions on all the other LLM's, they answered correctly and did not do what Cohere did.
What is going on here?
Thanks.
Brand new different browser, clean cache/etc.
Signed into Cohere Command A+
Started with a few basic questions, then asked a riddle. For 12 MINUTES it went into a feedback loop/crashed.
When it came back, I asked it "3 dog names" and it continued crashing.
Grocery Compare retweeted
The "world leading" AI platform Cohere Carney just bought into with $1.6B of our cash.
Cohere blocked me when I went to them about this.
I asked Cohere's latest Ai model a simple question: "What year was the Ford F-150 introduced" (Answer: 1975).
Cohere chatbot went into a feedback loop/crashed and keep claiming it's 1979!
I've been testing this model for the past 2 days, it keeps crashing.
Grocery Compare retweeted
Again: Asking BASIC QUESTIONS to Cohere, it KEEPS CRASHING.
I asked Cohere a simple question "What year was the Ford F-150 introduced" (Year was 1975)
Cohere went into a feedback loop claiming it was 1979!
We're supposed to trust Cohere's flagship Ai for medical/military/banking/etc???
Grocery Compare retweeted
Hi Vishveshwar:
I want to be clear: I want Canada to succeed. However after I've found a major glitch in their Ai:
Cohere instead blocked me and has yet to even address this critical flaw. It's been 2+ days now.
Here again (I screen recorded it), I've been asking Cohere's latest Ai model just BASIC QUESTIONS:
This was a NEW chat window.
- What's the population of Amsterdam?
- How many ounces in ml is in a can of Coke?
- Cheapest gas prices in Oshawa?
- Where do Yukon gold potatoes come from?
When finally I asked "What year was the Ford F-150 introduced?" This again led to the Cohere Command+ model to go into a feedback loop/crash.
No other Ai model on the market does this!
November 1st, 1:21am.
I asked Cohere's latest Ai model: "1+1", it answered "2" (That's correct).
I then asked: "A photographer takes a picture of a man, develops it, and then hangs the man. How is this possible?"
and it (again) crashed.
