00:00these are 15 new ai discoveries that prove the future is already here number 15 ai that can
00:09simulate entire worlds one of the biggest ai shifts right now is world modeling instead of
00:17just generating texts images or video ai is starting to build interactive environments that
00:24simulate how reality works that matters because these systems can train robots test decisions
00:32and predict outcomes before anything happens in the real world and companies are taking this
00:37seriously jan lukun's ami labs raised 500 million at a 3 billion valuation to build a world modeling
00:46platform google deepmind's genie 3 can generate interactive environments at 24 frames per second
00:52in 720p while nvidia's cosmos platform has already passed 2 million downloads number 14 living cells
01:02programmed like computers this is where ai starts pushing into biology in a seriously futuristic way
01:11researchers at tu darmstadt use deep learning and bayonese optimization to design a hybrid ribo switch
01:19that works like a boolean nand gate which is one of the core logic functions in computing the ai
01:26searched through thousands of possible rna sequences but scientists only had to test 82 variants in the
01:33lab to find the best design the same study also built a dual ribo switch that responds to two ligands
01:40letting gene expression switch on or off based on specific input combinations in other words living
01:47cells are starting to look programmable number 13 ai that acts like a real worker ai is no longer just
01:58helping with tasks here and there it's starting to function like an actual co-worker these systems can
02:05handle full desktop workflows write reports organize files and manage real business operations with minimal
02:13supervision and the numbers make that clear co-worker dot ai generated 2.9 million in revenue in 2025 with
02:22only 26 employees showing how lean ai first companies can become then there's junior from coos ai more than
02:312 000 companies paid deposits just to get a demo and junior reportedly writes 80 of the company's code
02:39starts half of all sales calls and manages 80 of communications number 12 ai solving problems humans
02:50couldn't crack this is where ai stops feeling like a tool and starts looking like a breakthrough machine
02:57deep minds alpha evolve found a way to multiply four by four matrices using only 48 scalar multiplications
03:06beating the long-standing record of 49 that may sound technical but it shows ai is beginning to solve math
03:14problems humans couldn't crack for years and it's not just theory alpha evolve also improved google's data
03:22center scheduling freeing up 0.7 of its global compute resources speeding up gemini training by 23
03:30percent cutting total training time by one percent and making the flash attention kernel 32.5 percent
03:38faster number 11 ai designing new medicines from scratch drug discovery is usually slow expensive and full
03:48of dead ends ai is starting to change that according to the outline ai can cut drug discovery time by
03:56around
03:5730 to 40 percent and reduce the cost per successful drug by about 2.8 billion dollars the market itself
04:06is expected to grow from 2.9 billion in 2025 to 13.4 billion by 2035 one of the clearest
04:14examples is in
04:16silico medicine its generative platform created 15 million virtual molecules but only 60 compounds had to be
04:25tested in the lab helping deliver a candidate to ind filling in just 18 months instead of the usual four
04:32to five years number 10 the rise of fully autonomous ai systems ai is starting to move beyond being a
04:42tool and
04:43into something more independent these new systems can plan execute and improve workflows with minimal
04:52supervision which is why so many companies are now testing them a 2025 survey found that 23 percent of
04:59organizations had already launched pilot programs using autonomous ai agents while 61 percent were still
05:07exploring deployment at the same time trust is becoming a real issue with public confidence falling from
05:1543 percent to 27 percent in just one year pantera capital also estimates u.s hyperscalers will spend over 650
05:26billion on ai in 2026 number nine supercomputers built just for ai discovery ai is now getting its own
05:37class of supercomputers built specifically for scientific discovery the u.s department of energy announced
05:44two ai focused supercomputers lux and discovery through a public private partnership worth more than a
05:52billion dollars these systems will use amd epyc processors and instinct gpus to speed up work in fusion
06:02materials discovery and national security and the bigger idea here is speed analysts project the ai
06:10supercomputer market will grow from 2.37 billion in 2025 to 2.86 billion by 2030 with systems like these
06:20expected to drive breakthroughs 10 to 100 times faster than human scientists number eight brain inspired
06:28chips that use almost no power one of the most exciting hardware breakthroughs in ai is neuromorphic computing
06:36where chips are designed to work more like the human brain instead of burning huge amounts of energy
06:42they solve problems with far more efficiency intel's loihai 2 chip delivers over a hundred times better
06:51energy efficiency than conventional cpus and gpus for optimization tasks while also giving sub-millisecond
07:00latency then there's the hollow point a system made from 1152 loihai 2 processors it simulates 1.15
07:11billion neurons supporting 128 billion synapses uses just 2600 watts and delivers 15 trillion 8-bit
07:21operations per second per watt number seven ai that works in teams not alone the next big
07:30step isn't just smarter ai it's ai systems working together like full teams these multi-agent setups can
07:39divide tasks coordinate with each other and keep projects moving over longer periods without one
07:46model doing everything alone a 2025 survey found that 88 of organizations already use ai in at least one
07:54function and 76 plan to deploy agentic ai within 12 months gartner predicts that 40 of enterprise
08:03applications will include task-specific agents by 2026 up from less than five percent today analysts also
08:12expect the agentic ai market to grow from 7.8 billion to around 52 billion by 2030 number six ai
08:21that
08:22understands the laws of physics some of the most powerful ai systems now work by understanding
08:29physics instead of just spotting patterns that makes them much more useful in areas like weather
08:35climate and fluid dynamics where real world laws matter huawei's pangu weather can generate global
08:4324-hour forecasts in 1.4 seconds on an nvidia v100 gpu which is about 10 000 times faster than
08:53traditional
08:54numerical methods deepmine's graphcast can produce 10-day forecasts in under a minute on a single tpu v4
09:03outperforming the operational hres system on more than 90 percent of 1380 test variables and reach a
09:1299.7 accuracy in the troposphere number five ai models that remember everything one of the biggest
09:22limits in ai has always been memory most models could sound smart for a few minutes then start losing the
09:30thread once the conversation or document got too long that's changing fast google's gemini 3 pro and
09:39meta's llama 4 scout now offer 10 million token context windows the largest advertised by any large
09:48language model open ai's gpt 4.1 supports 1 million tokens while anthropics clawed for sonnet handles 200
09:57000 tokens with less than five percent accuracy loss across the full context ai is starting to remember
10:05enough to handle massive tasks without constantly forgetting what came before number four ultra efficient
10:14ai that runs on less compute a lot of people assume better ai always means bigger models and more
10:22hardware but one of the most important breakthroughs is actually efficiency researchers at the university of
10:30texas and mit built a neuro symbolic system to solve the tower of hanoi with 95 percent success while
10:38using just one percent of the training energy and five percent of the operating energy required by standard
10:46visual language action models it also cut training time from days to just 34 minutes then there's
10:53mistral's mixed trawl 8x7b which has 46.7 billion parameters but activates only 12.9 billion per token
11:04making inference about six times faster than llama 270b number three ai that learns with far less data for years
11:15one of ai's biggest weaknesses was how much data it needed these systems were powerful but they were
11:23also hungry now that's starting to change on many image net benchmarks few shot models can reach 68 to 81
11:32accuracy using just five examples per class and methods like simple shot hit 81.5 without any meta training
11:42mit's compress m takes this even further by compressing state space models during training the compressed
11:49versions keep nearly the same accuracy while training up to 1.5 times faster in one case a 128 dimensional
11:58model was compressed to around 12 dimensions and still stayed competitive with about four times faster training
12:06number two ai running real businesses in the background ai is no longer just helping companies from the sidelines
12:15in some cases in some cases it's already running major parts of business one example is ven hub's fully autonomous
12:23smart store which opened at los angeles union station in november 2025 the store runs 24 7 carries more than
12:31400
12:32products and serves a location with over 60 000 daily passengers it was also installed and activated in just a
12:39few days
12:40then there's a widely shared experiment where entrepreneur ravindu himansha let an ai agent run his saas
12:49business for 30 days the system closed a 12 000 annual contract and increased monthly revenue from 32 000 to
12:5845 000
13:00number one ai accelerating scientific discovery at scale this may be the clearest sign that the future
13:10is already here ai is now accelerating scientific discovery at a speed no human team could match
13:18alone in 2022 deep mind released alpha fold 2 predictions for over 200 million protein structures achieving in a
13:27single year what would have taken hundreds of millions of years to do experimentally since then the open
13:34access database has been used by more than 3 million researchers across 190 countries and the impact goes
13:41beyond scale one independent analysis found that researchers using alpha fold 2 discovered over 40 percent more
13:51novel protein structures and alpha fold linked studies were twice as likely to be cited by patents compared with
13:59typical structural biology papers if you made it this far let us know what you think in the comment section
14:05below
Comments