Nifal Adam
@nifaladam.bsky.social
Super Affiliate - Over $3M in ad spend 🧢
Indiehacking Sales - $ XX,XXX 🔥
IIT grad 🎓
startupbolt.com
Sharing AI trends & SaaS tips 4x/week | Follow for insights.
✨ NextJS Website Builder 👉 https://www.startupbolt.com/
Indiehacking Sales - $ XX,XXX 🔥
IIT grad 🎓
startupbolt.com
Sharing AI trends & SaaS tips 4x/week | Follow for insights.
✨ NextJS Website Builder 👉 https://www.startupbolt.com/
Stay tuned for Day 3 updates tomorrow! 🚀
Follow me to get the latest updates first!
#AI #ArtificialIntelligence #OpenAI #TechUpdates
Follow me to get the latest updates first!
#AI #ArtificialIntelligence #OpenAI #TechUpdates
December 6, 2024 at 7:41 PM
Stay tuned for Day 3 updates tomorrow! 🚀
Follow me to get the latest updates first!
#AI #ArtificialIntelligence #OpenAI #TechUpdates
Follow me to get the latest updates first!
#AI #ArtificialIntelligence #OpenAI #TechUpdates
Why it matters:
• Democratizes advanced AI training
• Enables domain-specific expertise
• Uses minimal training data
• Maintains speed of smaller models
• Cost-effective solutions
• Democratizes advanced AI training
• Enables domain-specific expertise
• Uses minimal training data
• Maintains speed of smaller models
• Cost-effective solutions
December 6, 2024 at 7:41 PM
Why it matters:
• Democratizes advanced AI training
• Enables domain-specific expertise
• Uses minimal training data
• Maintains speed of smaller models
• Cost-effective solutions
• Democratizes advanced AI training
• Enables domain-specific expertise
• Uses minimal training data
• Maintains speed of smaller models
• Cost-effective solutions
Beyond healthcare:
Successfully tested in:
• Biochemistry
• AI safety
• Legal
• Finance
• Engineering
• Insurance
Successfully tested in:
• Biochemistry
• AI safety
• Legal
• Finance
• Engineering
• Insurance
December 6, 2024 at 7:41 PM
Beyond healthcare:
Successfully tested in:
• Biochemistry
• AI safety
• Legal
• Finance
• Engineering
• Insurance
Successfully tested in:
• Biochemistry
• AI safety
• Legal
• Finance
• Engineering
• Insurance
Early access program:
• Open to universities, researchers, enterprises
• Focus on complex tasks requiring expert teams
• Already used by Thomson Reuters for legal AI
• Applications open
• Open to universities, researchers, enterprises
• Focus on complex tasks requiring expert teams
• Already used by Thomson Reuters for legal AI
• Applications open
December 6, 2024 at 7:41 PM
Early access program:
• Open to universities, researchers, enterprises
• Focus on complex tasks requiring expert teams
• Already used by Thomson Reuters for legal AI
• Applications open
• Open to universities, researchers, enterprises
• Focus on complex tasks requiring expert teams
• Already used by Thomson Reuters for legal AI
• Applications open
Example case study:
Patient symptoms → Gene prediction
• Input: Medical history + symptoms
• Output: Ranked gene list + reasoning
• Grading: Higher scores for correct genes ranked higher
• Success: Found TSC2 as top candidate
Patient symptoms → Gene prediction
• Input: Medical history + symptoms
• Output: Ranked gene list + reasoning
• Grading: Higher scores for correct genes ranked higher
• Success: Found TSC2 as top candidate
December 6, 2024 at 7:41 PM
Example case study:
Patient symptoms → Gene prediction
• Input: Medical history + symptoms
• Output: Ranked gene list + reasoning
• Grading: Higher scores for correct genes ranked higher
• Success: Found TSC2 as top candidate
Patient symptoms → Gene prediction
• Input: Medical history + symptoms
• Output: Ranked gene list + reasoning
• Grading: Higher scores for correct genes ranked higher
• Success: Found TSC2 as top candidate
How RFT works:
• Upload training data (JSONL format)
• Define grading criteria (0-1 scale)
• Model learns through reinforcement
• Validation on separate dataset
• Performance tracking via dashboard
• Upload training data (JSONL format)
• Define grading criteria (0-1 scale)
• Model learns through reinforcement
• Validation on separate dataset
• Performance tracking via dashboard
December 6, 2024 at 7:41 PM
How RFT works:
• Upload training data (JSONL format)
• Define grading criteria (0-1 scale)
• Model learns through reinforcement
• Validation on separate dataset
• Performance tracking via dashboard
• Upload training data (JSONL format)
• Define grading criteria (0-1 scale)
• Model learns through reinforcement
• Validation on separate dataset
• Performance tracking via dashboard
Results from testing:
O1-mini with RFT outperformed base O1:
• Top@1: 31% vs 25%
• Top@5: 41% vs 38%
• Top@max: 45% vs 38%
All on completely new genes (no overlap in training)
O1-mini with RFT outperformed base O1:
• Top@1: 31% vs 25%
• Top@5: 41% vs 38%
• Top@max: 45% vs 38%
All on completely new genes (no overlap in training)
December 6, 2024 at 7:41 PM
Results from testing:
O1-mini with RFT outperformed base O1:
• Top@1: 31% vs 25%
• Top@5: 41% vs 38%
• Top@max: 45% vs 38%
All on completely new genes (no overlap in training)
O1-mini with RFT outperformed base O1:
• Top@1: 31% vs 25%
• Top@5: 41% vs 38%
• Top@max: 45% vs 38%
All on completely new genes (no overlap in training)
Real-world example: Rare disease diagnosis
• Partnered with Berkeley Lab & Charité Hospital
• Model analyzes patient symptoms to identify causative genes
• Addresses 300M+ people globally affected by rare diseases
• Reduces diagnostic time from months/years
• Partnered with Berkeley Lab & Charité Hospital
• Model analyzes patient symptoms to identify causative genes
• Addresses 300M+ people globally affected by rare diseases
• Reduces diagnostic time from months/years
December 6, 2024 at 7:41 PM
Real-world example: Rare disease diagnosis
• Partnered with Berkeley Lab & Charité Hospital
• Model analyzes patient symptoms to identify causative genes
• Addresses 300M+ people globally affected by rare diseases
• Reduces diagnostic time from months/years
• Partnered with Berkeley Lab & Charité Hospital
• Model analyzes patient symptoms to identify causative genes
• Addresses 300M+ people globally affected by rare diseases
• Reduces diagnostic time from months/years
Key features of RFT:
• Uses same RL algorithms that improved O1
• Requires only dozens of examples
• Allows models to learn new reasoning patterns
• Launching publicly in early 2024
• Uses same RL algorithms that improved O1
• Requires only dozens of examples
• Allows models to learn new reasoning patterns
• Launching publicly in early 2024
December 6, 2024 at 7:41 PM
Key features of RFT:
• Uses same RL algorithms that improved O1
• Requires only dozens of examples
• Allows models to learn new reasoning patterns
• Launching publicly in early 2024
• Uses same RL algorithms that improved O1
• Requires only dozens of examples
• Allows models to learn new reasoning patterns
• Launching publicly in early 2024
Stay tuned for Day 2 updates tomorrow! 🚀
Follow me to get the latest updates first!
#AI #ArtificialIntelligence #OpenAI #TechUpdates
Follow me to get the latest updates first!
#AI #ArtificialIntelligence #OpenAI #TechUpdates
December 6, 2024 at 5:42 PM
Stay tuned for Day 2 updates tomorrow! 🚀
Follow me to get the latest updates first!
#AI #ArtificialIntelligence #OpenAI #TechUpdates
Follow me to get the latest updates first!
#AI #ArtificialIntelligence #OpenAI #TechUpdates
Coming soon with o1:
• Enhanced capability for more compute-intensive tasks
• Web browsing
• File uploads
• API access with new features like image understanding
• Enhanced capability for more compute-intensive tasks
• Web browsing
• File uploads
• API access with new features like image understanding
December 6, 2024 at 5:42 PM
Coming soon with o1:
• Enhanced capability for more compute-intensive tasks
• Web browsing
• File uploads
• API access with new features like image understanding
• Enhanced capability for more compute-intensive tasks
• Web browsing
• File uploads
• API access with new features like image understanding
Demo 3: o1 Pro mode
Solved challenging chemistry protein problem:
• Complex multi-criteria analysis
• Correctly identified retinochisen
• Demonstrated deep domain knowledge
• Solved in 53 seconds
Solved challenging chemistry protein problem:
• Complex multi-criteria analysis
• Correctly identified retinochisen
• Demonstrated deep domain knowledge
• Solved in 53 seconds
December 6, 2024 at 5:42 PM
Demo 3: o1 Pro mode
Solved challenging chemistry protein problem:
• Complex multi-criteria analysis
• Correctly identified retinochisen
• Demonstrated deep domain knowledge
• Solved in 53 seconds
Solved challenging chemistry protein problem:
• Complex multi-criteria analysis
• Correctly identified retinochisen
• Demonstrated deep domain knowledge
• Solved in 53 seconds
Demo 2: Multimodal capabilities
Solved complex space datacenter cooling problem:
• Analyzed hand-drawn diagram
• Calculated radiator surface area (2.42M sq meters)
• Handled ambiguous parameters intelligently
• State-of-the-art performance on MMMU & MathVista benchmarks
Solved complex space datacenter cooling problem:
• Analyzed hand-drawn diagram
• Calculated radiator surface area (2.42M sq meters)
• Handled ambiguous parameters intelligently
• State-of-the-art performance on MMMU & MathVista benchmarks
December 6, 2024 at 5:42 PM
Demo 2: Multimodal capabilities
Solved complex space datacenter cooling problem:
• Analyzed hand-drawn diagram
• Calculated radiator surface area (2.42M sq meters)
• Handled ambiguous parameters intelligently
• State-of-the-art performance on MMMU & MathVista benchmarks
Solved complex space datacenter cooling problem:
• Analyzed hand-drawn diagram
• Calculated radiator surface area (2.42M sq meters)
• Handled ambiguous parameters intelligently
• State-of-the-art performance on MMMU & MathVista benchmarks
Demo 1: Speed improvement
Sample history query about Roman emperors:
• o1: 14 seconds
• o1 Preview: 33 seconds
Plus higher accuracy on complex historical details
Sample history query about Roman emperors:
• o1: 14 seconds
• o1 Preview: 33 seconds
Plus higher accuracy on complex historical details
December 6, 2024 at 5:42 PM
Demo 1: Speed improvement
Sample history query about Roman emperors:
• o1: 14 seconds
• o1 Preview: 33 seconds
Plus higher accuracy on complex historical details
Sample history query about Roman emperors:
• o1: 14 seconds
• o1 Preview: 33 seconds
Plus higher accuracy on complex historical details
ChatGPT Pro features:
• Unlimited access to models (o1, GPT-4)
• Advanced voice mode
• o1 Pro mode for enhanced performance
• Higher reliability on complex tasks
• Will get additional features during the 12 days
• Unlimited access to models (o1, GPT-4)
• Advanced voice mode
• o1 Pro mode for enhanced performance
• Higher reliability on complex tasks
• Will get additional features during the 12 days
December 6, 2024 at 5:42 PM
ChatGPT Pro features:
• Unlimited access to models (o1, GPT-4)
• Advanced voice mode
• o1 Pro mode for enhanced performance
• Higher reliability on complex tasks
• Will get additional features during the 12 days
• Unlimited access to models (o1, GPT-4)
• Advanced voice mode
• o1 Pro mode for enhanced performance
• Higher reliability on complex tasks
• Will get additional features during the 12 days
o1 improvements:
• 34% fewer major mistakes
• 50% faster response times
• New multimodal capabilities (image + text)
• Better instruction following
• Adaptive thinking (fast for simple queries, thorough for complex ones)
• 34% fewer major mistakes
• 50% faster response times
• New multimodal capabilities (image + text)
• Better instruction following
• Adaptive thinking (fast for simple queries, thorough for complex ones)
December 6, 2024 at 5:42 PM
o1 improvements:
• 34% fewer major mistakes
• 50% faster response times
• New multimodal capabilities (image + text)
• Better instruction following
• Adaptive thinking (fast for simple queries, thorough for complex ones)
• 34% fewer major mistakes
• 50% faster response times
• New multimodal capabilities (image + text)
• Better instruction following
• Adaptive thinking (fast for simple queries, thorough for complex ones)
o1 Performance Gains vs GPT-4:
• Competition Math: 13.4% → 83.3%
• Competition Code: 11% → 89%
• PhD Science Questions: 56.1% → 78%
• Matching/exceeding expert human performance (69.7%)
• Competition Math: 13.4% → 83.3%
• Competition Code: 11% → 89%
• PhD Science Questions: 56.1% → 78%
• Matching/exceeding expert human performance (69.7%)
December 6, 2024 at 5:42 PM
o1 Performance Gains vs GPT-4:
• Competition Math: 13.4% → 83.3%
• Competition Code: 11% → 89%
• PhD Science Questions: 56.1% → 78%
• Matching/exceeding expert human performance (69.7%)
• Competition Math: 13.4% → 83.3%
• Competition Code: 11% → 89%
• PhD Science Questions: 56.1% → 78%
• Matching/exceeding expert human performance (69.7%)
lol probably true. That’s probably true. The sudden automatic braking caused my cousin to get rear-ended in a car crash. It took nearly two months to get it fixed. This is UAE though.
December 6, 2024 at 2:14 PM
lol probably true. That’s probably true. The sudden automatic braking caused my cousin to get rear-ended in a car crash. It took nearly two months to get it fixed. This is UAE though.
Its been 10 years since I stopped caring about which day of the week it is (Ever since I started a business lol)
December 6, 2024 at 1:55 PM
Its been 10 years since I stopped caring about which day of the week it is (Ever since I started a business lol)
Is this true? The service and repairs take waaaaay too long, but its been a pretty smooth and safe car.
December 6, 2024 at 1:54 PM
Is this true? The service and repairs take waaaaay too long, but its been a pretty smooth and safe car.
Amazon is rewriting the AI playbook, and I’m here for it.
Follow me for more breakdowns like this! 💡
Once I hit 1,000 followers, I’ll share the Ultimate Marketing Sheet for organic SaaS sales with all my followers. 🛠️
#AI #Amazon #IndieHacking
Follow me for more breakdowns like this! 💡
Once I hit 1,000 followers, I’ll share the Ultimate Marketing Sheet for organic SaaS sales with all my followers. 🛠️
#AI #Amazon #IndieHacking
December 6, 2024 at 1:46 PM
Amazon is rewriting the AI playbook, and I’m here for it.
Follow me for more breakdowns like this! 💡
Once I hit 1,000 followers, I’ll share the Ultimate Marketing Sheet for organic SaaS sales with all my followers. 🛠️
#AI #Amazon #IndieHacking
Follow me for more breakdowns like this! 💡
Once I hit 1,000 followers, I’ll share the Ultimate Marketing Sheet for organic SaaS sales with all my followers. 🛠️
#AI #Amazon #IndieHacking
📈 What’s Next?
Amazon plans to:
• Invest billions in AI hardware & models
• Make AI affordable for EVERY business
• Continue shaping the future of AI innovation
This is just the beginning of Amazon’s AI ambitions.
Amazon plans to:
• Invest billions in AI hardware & models
• Make AI affordable for EVERY business
• Continue shaping the future of AI innovation
This is just the beginning of Amazon’s AI ambitions.
December 6, 2024 at 1:46 PM
📈 What’s Next?
Amazon plans to:
• Invest billions in AI hardware & models
• Make AI affordable for EVERY business
• Continue shaping the future of AI innovation
This is just the beginning of Amazon’s AI ambitions.
Amazon plans to:
• Invest billions in AI hardware & models
• Make AI affordable for EVERY business
• Continue shaping the future of AI innovation
This is just the beginning of Amazon’s AI ambitions.
💡 Tackling AI Risks
Amazon’s approach to AI hallucinations:
• RefChecker: Detects hallucinated outputs.
• Amazon Q: Provides source citations for transparency.
They’re not just building AI—they’re making it trustworthy.
Amazon’s approach to AI hallucinations:
• RefChecker: Detects hallucinated outputs.
• Amazon Q: Provides source citations for transparency.
They’re not just building AI—they’re making it trustworthy.
December 6, 2024 at 1:46 PM
💡 Tackling AI Risks
Amazon’s approach to AI hallucinations:
• RefChecker: Detects hallucinated outputs.
• Amazon Q: Provides source citations for transparency.
They’re not just building AI—they’re making it trustworthy.
Amazon’s approach to AI hallucinations:
• RefChecker: Detects hallucinated outputs.
• Amazon Q: Provides source citations for transparency.
They’re not just building AI—they’re making it trustworthy.