Cureus innovator

MD. MPH Yale alumni,🧑‍🎓 Physician Scientist. Public health. #ManchesterUnitedFan

[Cureus innovator @JananthanMD]

AI is often trained to win human approval, not to survive contact with reality. That works for tone and usability. It becomes dangerous when we ask models to solve problems no human can easily verify. https://t.co/2ptSpDPEp9

[Cureus innovator @JananthanMD]

@ReliableAIDeliv Agent Separation of duties fixes authorization, but an independent LLM reviewer is still bounded by RLHF heuristics and reward hacking. It has to be fixed at the foundation level .

[Cureus innovator @JananthanMD]

AI is often trained to win human approval, not to survive contact with reality. That works for tone and usability. It becomes dangerous when we ask models to solve problems no human can easily verify. https://t.co/2ptSpDPEp9

[Cureus innovator @JananthanMD]

@anishgiri It’s not the model , it’s RHLF

[Aaron Burnett @aaronburnett]

Thought provoking. One core claim here that I find intriguing is that (slight paraphrase) “allocating more compute to inference suggests you aren’t growing fast because if you were you’d put more compute toward training to get a better model”

This makes sense on the surface but it strikes me as something only valid in a compute constrained paradigm which can be local not global.

For example Spacex is guiding toward using only ~10% of their compute for training. You can argue that means it’s because Spacex isn’t leading (but wouldn’t they want to catch up faster then) or you can argue that majority of compute dedicated to training isn’t a default true assumption.

Compute co-design also suggests that efficiency in training would dramatically reduce the required compute for training.

Leaves the question why wouldn’t SpaceX just dedicated all their compute to train the god model? 🤔

[Prime Intellect @PrimeIntellect]

Prime Agent is built on pi and fully open-source with an open license

https://t.co/dN4YQ3IPdt

[Nicolas Camara @nickscamara_]

introducing anydoc

now your agents get 100x faster local parsing for pdf, docx, pptx & 10 more formats

- sub-5ms md conversion
- 500 docx files in 1.7s
- top quality across all 13 formats
- rust-based
- open source

already powering @firecrawl /parse

https://t.co/fVsRsYFNyC

[Google for Developers @googledevs]

No more hardcoded endpoints. Model routing for API Gateway is in Public Preview! Access Gemini, Claude, and OSS models through 1 single endpoint using the model name as a parameter in your OpenAPI spec.

Read the guide to configure yours ➔ https://t.co/K71ZvP0uD5

[Kevin Jiang @jiang_kev]

We just raised $5.7M for @PolarBrowser, the AI browser that beats Anthropic and OpenAI on every major web agent benchmark.

- 4,500,000+ actions taken for users, automating sales, recruiting, and ops
- One company cancelled Clay and saves 25+ hrs/wk per person
- Team is from MIT, YC, Prod, Citadel, Jane Street, Perplexity, Modal, and Apple

100 hours of work. From 30 seconds of typing.

Download the world's most powerful AI browser: https://t.co/lUOm01Stm0

[Cureus innovator @JananthanMD]

@ln1nited Piastri should be a great driver and leclerc should be a good driver

[Cureus innovator @JananthanMD]

@ssi @OpenAI and @claudeai will be the Nokia ‘s of AI world and @ssi will be the iPhone era in AI

[codila @0xCodila]

Google just released free 1-hour course on building agentic knowledge Graphs from 0% to 100%:

10% → 4:01 - how to build a GraphRAG agent
30% → 15:00 - Graph Engineering explanation
55% → 30:00 - Agentic search Engineering
80% → 35:48 - Graph Engineering practice
100% → 47:06 - self-improving agents in graphs

this free Google course mass replaces a $500 graph engineering bootcamp - learn it in 60 min to 100%

watch it today - then read the full graph playbook in the article below ↓

[Dimple Arora @DimpleAroro23s]

POV: using Claude Fable 5 to rename variables

[NIH Innovates @NIH_Innovates]

📢 New Funding Opportunity from the NIH Common Fund’s PRIMED-AI program: Researchers are invited to apply for support to help move validated AI-powered clinical decision support tools from prototype to real-world clinical implementation.

📅 Applications due: October 19, 2026
➡️ Learn more: https://t.co/mJ5VHGrVBd

[Cureus innovator @JananthanMD]

@HomerPavlos That’s also a problem, human beings will be hooked in , without creativity of their own .

[NIK @ns123abc]

🚨The White House announces $5 BILLION+ for the Genesis Mission

>15+ agencies (DOE, NIH, NSF, NASA, Department of War, DHS, USDA, EPA…)
>16 challenges (pediatric cancer, grid scaling, microelectronics, weapons component design, nuclear signature attribution, autonomous labs…)

Awardees get the American Science Cloud

>Microsoft commits $40M Azure credits + $20M in engineering services
>Google commits $40 million in AI tokens + cloud credits and DeepMind portfolio + 1 free year of Gemini for all DOE researchers

DOE: "the largest response to a funding opportunity in DOE history."

AI for science, its happening

[Cureus innovator @JananthanMD]

@adocomplete I already canceled the subscription I had for more than a year, it’s so stupid with your restrictions with biology and usage limits . Done with your BS.

[Cureus innovator @JananthanMD]

@TheLaurenChen If you care about cinema, I would say I avoid this tweet .

[Lumina @LuminaBench]

🚨 New Gemini 3.5 Pro leak just surfaced

Google appears to be preparing its next flagship Gemini model:

• Internal codename reportedly: Cappuccino
• Expected to introduce a new Deep Thinking mode
• Rumoured 2M-token context window
• Stronger tool use, planning and multi-step execution
• Release reportedly targeted for later this month still after delays
• Major gains in reasoning, coding and long-horizon agent tasks
• Alleged benchmark improvements across web development, SVG generation and complex problem-solving

Some leaked builds may already be appearing through LMArena and Google’s Antigravity platform, but the model identity and benchmark claims remain unconfirmed.

Are you looking forward to Gemini 3.5 Pro being released?

[Codez @0xCodez]

Andrej Karpathy just dropped a 6-hour course on how to build LLMs from scratch:

• 00:00 - Deep dive into LLMs like ChatGPT
• 03:31:23 - Building ChatGPT from scratch in live
• 05:27:43 - How to use LLMs (Karpathy method)

This course will replace a $90K Stanford LLM master’s degree.

Start watching today, then read how to become an AI engineer in article below.