Joined August 2020
- Tweets 833
- Following 301
- Followers 711
- Likes 739
171 Photos and videos
Jun 14
Week one since flash-1-mini launch.
Recap:
→ 4B model live on Hugging Face (Apache 2.0)
→ Public benchmark methodology published
→ Built on Canadian sovereign compute (Telus)
→ flash-1 (9B) arrives September 30
→ flash-1-pro (27B MoE) March 31, 2027
The flash series is just getting started.
2
119
Jun 13
We were building a Canadian legal AI model. We stopped.
Our own model invented 7 of 10 legal citations.
Even Harvey runs retrieval - not trained-in case law.
flash-1 is now a foundation model for Canadian work.
Citations get retrieved from licensed sources.
We accepted RAG.
1
82
Jun 13
What shipped in the last week:
→ flash-1-mini (4B model, Apache 2.0)
→ CBLRE evaluation set
→ Canadian AI evaluation methodology v1.0
→ Model benchmarking methodology v1.0
All open. All reproducible. All on the SimpleDirect blog.
54
Jun 12
Confidentiality changes the math.
For Big Law associates,
M&A analysts,
in-house counsel
- client work doesn't go into ChatGPT.
Self-hosted specialized AI changes that calculation.
Same hardware.
Same data.
Your model.
Specialized for the work.
58
Jun 12
What 'AI you can own' actually means:
→ Apache 2.0 weights (download forever)
→ Open RAG infrastructure
→ Self-deploy on your hardware
→ Fine-tune on your proprietary corpus
→ Resulting weights are yours
→ Citation grounding you can audit
The whole stack, not just the model file.
49
Jun 12
Open weights and open systems are not the same thing.
A vendor can ship Apache 2.0 weights and lock the runtime.
The license governs the model file.
The license does not govern the user experience.
If you can't swap the model, it's not really open.
The default path is the whole game.
28
Jun 10
Most AI companies show you a score and hide the test.
We're doing the opposite.
Today we published CBLRE -
a test for Canadian legal AI,
written and checked by legal experts,
in both official languages.
The test first.
The scores later - including our own.
Because a score you can't check isn't worth much.
1
63
Jun 10
Un modèle peut obtenir 90 % au test MMLU de droit professionnel et quand même :
→ Mal appliquer le Code civil du Québec
→ Inventer des citations qui semblent crédibles
→ Être plus faible en français qu'en anglais
→ Refuser des questions légitimes tout en répondant à celles qu'il devrait refuser
CBLRE mesure les quatre.
54
Jun 10
A model can score 90% on MMLU professional law and still:
→ Misapply the Civil Code of Québec
→ Fabricate citations that look real
→ Score weaker in French than in English
→ Refuse safe questions while answering ones it shouldn't
CBLRE measures all four.
23
Jun 10
Les référentiels juridiques américains couvrent la common law, en anglais seulement.
Les référentiels multilingues généraux testent le français courant, pas le français juridique.
Et aucun ne touche au droit civil québécois - une tradition juridique distincte.
CBLRE mesure ce qui compte vraiment pour le travail réglementé au Canada.
19
Jun 10
US legal benchmarks are common-law and English-only.
General multilingual benchmarks test French fluency, not legal French.
And none of them touch Quebec civil law - a different legal tradition entirely.
CBLRE measures what actually matters for Canadian regulated work.
15
Jun 9
A lawyer in Montreal asks a legal AI a question in French.
It answers worse than it would in English.
But on average, the model still 'scores well.'
That average is hiding a failure.
CBLRE scores each language on its own - so it can't.
1
25
Jun 9
CBLRE tracks:
→ Common law (EN)
→ Quebec civil law (FR)
→ Constitutional / Charter
→ Privacy compliance (EN/FR paired)
→ Citation integrity
→ Safety calibration
Each scored independently.
No blended headline number.
25
Jun 9
flash-1-mini is fine-tuned on a Canadian bilingual legal and regulatory corpus.
No public benchmark existed to measure work like this.
So we built one.
CBLRE - the Canadian Bilingual Legal and Regulatory Evaluation.
Six tracks. Expert-reviewed. Bilingual ground truth.
getsimpledirect.com/en/news/…
2
36
Jun 8
The work you can't paste into ChatGPT.
That's what specialized open Canadian AI is for.
Models you can own.
On your hardware.
With your data.
Audit the weights.
Fine-tune on your corpus.
AI you can own.
1
33
Jun 8
Canada has been buying legal AI with no public way to measure it.
So we published CBLRE.
The first bilingual benchmark for Canadian legal and regulatory AI.
Six tracks.
Reproducible.
Bilingual ground truth.
Anyone can run it - including against us.
getsimpledirect.com/en/news/…
1
30
Jun 8
Last week we shipped flash-1-mini.
A 4B Canadian business, legal, and regulatory AI model.
Runs on a MacBook.
Apache 2.0.
Open weights.
The first of the flash series.
huggingface.co/SimpleDirect/…
1
3
480
SimpleDirect retweeted
May 25
Toronto Tech Week.
Caught Minister @EvanLSolomon on the sideline.
For a year I've said it here:
A G7 country shouldn't have to rent its own compute.
Canada's finally moving.
Rimouski sold out. Vancouver next.
Real GPUs. Real money. Canadian soil.
Early. But real.
Good to show up for the people building it.
2
1
21
1,234
SimpleDirect retweeted
May 15
San Francisco peaked before ChatGPT.
Most people think the AI boom made the Valley.
The data says it ended it.
Here's what changed - and why Microsoft's 70-rule three weeks ago is the proof, not the cause:
8
3
33
2,852
SimpleDirect retweeted
May 4
OpenAI raised at $852B.
Secondary market: $880B.
A 3% premium.
Anthropic raised at $380B.
Secondary market: $1T.
A 165% premium.
Same market. Same buyers.
The market is pricing AI capex as a liability now.
The bigger you are, the more they're discounting you.
6
2
12
1,294