Мир сегодня с "Юрий Подоляка"

Труха⚡️Україна

Паглядзець

Николаевский Ванёк

Паглядзець

Мир сегодня с "Юрий Подоляка"

Паглядзець

Труха⚡️Україна

Паглядзець

Николаевский Ванёк

Паглядзець

Axis of Ordinary

Memetic and cognitive hazards.
Substack: https://axisofordinary.substack.com/

Рэйтынг TGlist

ТыпПублічны

Вертыфікацыя

Не вертыфікаваны

Надзейнасць

Не надзейны

РазмяшчэннеУзбекістан

МоваІншая

Дата стварэння каналаOct 19, 2020

Дадана ў TGlist

May 22, 2024

Я ўладальнік канала

Гісторыя змяненняў

Прыкрепленая група

Extraordinary Comments

Статыстыка Тэлеграм-канала Axis of Ordinary

Падрабязней

Падпісчыкаў

3 744

24 гадз.

2-0.1%Тыдзень

10-0.3%Месяц

40.1%

Індэкс цытавання

0

Згадкі0Рэпостаў на каналах0Згадкі на каналах0

Сярэдняе ахоп 1 паста

597

12 гадз.195

0.5%24 гадз.597

207.9%48 гадз.606

145.3%

Узаемадзеянне (ER)

3.61%

Рэпостаў11Каментары6Рэакцыі3

Узаемадзеянне па ахопу (ERR)

5.18%

24 гадз.0%Тыдзень

1.86%Месяц

0.48%

Ахоп 1 рэкламнага паста

597

1 гадз.10918.26%1 – 4 гадз.11619.43%4 - 24 гадз.58698.16%

Падрабязней

Падключыце нашага бота да канала і даведайцеся пол аўдыторыі гэтага канала.

Усяго пастоў за 24 гадзіны

8

Дынаміка

Апошнія публікацыі ў групе "Axis of Ordinary"

Усе пасты

23.04.202519:26

Links for 2025-04-23

AI

1. OpenAI’s o3 now outperforms 94% of expert virologists. https://www.ai-frontiers.org/articles/ais-are-disseminating-expert-level-virology-skills

2. TTRL: Test-Time Reinforcement Learning — A novel method for training LLMs using RL on *unlabeled* data by utilizing the priors in the pre-trained models. https://github.com/PRIME-RL/TTRL

3. Learning Adaptive Parallel Reasoning with Language Models https://arxiv.org/abs/2504.15466

4. Dynamic Early Exit in Reasoning Models https://arxiv.org/abs/2504.15895

5. RAGEN: Understanding Self-Evolution in LLM Agents via Multi-Turn Reinforcement Learning https://ragen-ai.github.io/

6. Anthropic believes that fully AI employees will soon start working in companies https://www.axios.com/2025/04/22/ai-anthropic-virtual-employees-security

7. Analyzing o3 and o4-mini with ARC-AGI. o4-mini: o1-pro-level performance, 220 × cheaper https://arcprize.org/blog/analyzing-o3-with-arc-agi

8. A father’s quest for diagnosis inspired a disruptive AI solution https://news.microsoft.com/source/emea/features/a-fathers-quest-for-diagnosis-inspired-a-disruptive-ai-solution/

9. “Our results suggest that models are rapidly improving, and the best frontier models are held back by only a few key subcapabilities.” [PDF] https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/6807879ce7b1b5f5163f4a32_RepliBenchPaper.pdf

10. 2 Big Questions for AI Progress in 2025-2026 https://helentoner.substack.com/p/2-big-questions-for-ai-progress-in

11. All AI datacenters are vulnerable to Chinese espionage, new report says https://time.com/7279123/ai-datacenter-superintelligence-china-trump-report/

12. AI supercomputers double in performance every 9 months, cost billions of dollars, and require as much power as mid-sized cities. Companies now own 80% of all AI supercomputers, while governments’ share has declined. https://epoch.ai/blog/trends-in-ai-supercomputers

13. Corrupted by Reasoning: Reasoning Language Models Become Free-Riders in Public Goods Games https://www.lesswrong.com/posts/M6dXdCbdoLSpHt8v3/corrupted-by-reasoning-reasoning-language-models-become-free

14. LLMs are Greedy Agents: Effects of RL Fine-tuning on Decision-Making Abilities https://arxiv.org/abs/2504.16078

Miscellaneous

1. Interpreting the retinal neural code for natural scenes: From computations to neurons https://www.cell.com/neuron/fulltext/S0896-6273(23)00467-1

2. Towards circuit mechanisms of the creative process: Describing the functions, mechanisms and neural correlates of creativity https://osf.io/preprints/psyarxiv/mpbgy_v1

3. How prediction markets create harmful outcomes: a case study https://bobjacobs.substack.com/p/how-prediction-markets-can-create

23.04.202516:34

"You Don't Have the Cards."

23.04.202515:53

Of course, these models still perform poorly in research-level mathematics, and such results are the exception. But that will soon change. These are just the first sparks (move 37) of the realm beyond.

https://x.com/bremen79/status/1914738549276205241

23.04.202515:18

A Practical Guide to Superhuman Mathematical AI
(written by o3)

1. Data Preparation
1.1. Download and clean formal libraries (Lean mathlib, Isabelle AFP, Coq, Metamath).
1.2. Crawl/tokenize informal sources (arXiv LaTeX, MathOverflow).
1.3. Seed with a small verified synthetic proof–theorem set.

2. Pre-train FM-Math
2.1. Choose a 10–100 B-param transformer with syntax-aware tokenization.
2.2. Train on formal + informal + seed data.

3. Graph-Based Encoder & Retrieval
3.1. Represent proof state as a graph (goals, hypotheses, lemma nodes).
3.2. Wrap FM-Math in a Graph Transformer over that graph.
3.3. Index lemma embeddings in a vector DB for k-NN retrieval.

4. RL Environment with Proof Assistants
4.1. Embed Lean/Isabelle as step(tactic) → (new state or error).
4.2. Reward +1 for completed proofs, shaped bonuses for closed subgoals, penalties for failures.

5. Dual-Head Agent
5.1. Conjecturer: mines gaps & analogies in the library graph.
5.2. Prover: outputs tactic/lemma distributions given current graph-state.

6. Search Orchestrator
6.1. Async MCTS (AlphaZero style), policy/value from dual-head.
6.2. Prune branches on verifier failure; share nodes across trees (SKEST).

7. Synthetic-Data Factory
7.1. Forward enumeration: random axiom subsets → depth-d consequences.
7.2. Mutation: weaken/strengthen or permute existing theorems.
7.3. Batch-run LeanNavigator agents to generate millions of proofs.
7.4. Filter via “TP-as-a-Judge” (batch verification).
7.5. Add only verified items back to replay buffers.

8. RL + Self-Play Loop
8.1. Initialize replay buffer with human & synthetic proofs.
8.2. Alternate:
(a) Off-policy RL (PPO) on buffer.
(b) Self-play jobs generating new proofs via the orchestrator.
8.3. Include an intrinsic-curiosity bonus for novel proof-state embeddings.

9. Phase 0: Instrumentation (2024–mid-25)
9.1. Validate FM-Math on FrontierMath, miniF2F, ProofNet.
9.2. Ensure end-to-end FM → graph → prover → verifier at scale.
9.3. Publish baselines.

10. Phase 1: Neuro-Symbolic MVP (mid-25–26)
10.1. ≥50 % FrontierMath; ≥80 % miniF2F.
10.2. Autoformalize first-year undergrad algebra texts at ≥70 %.
10.3. Ablate graph vs. flat encoding and RL+MCTS vs. pure policy.

11. Phase 2: Scaling & Data Gen (26–28)
11.1. Run LeanNavigator++ at 1 K GPUs → ≥1 B tokens verified.
11.2. Autonomously prove ≥40 % of theorems in a graduate corpus.
11.3. Introduce hierarchical proof planning: outline → tactics.

12. Phase 3: Open-Problem Sprint (28–31)
12.1. Focus AI+human teams on one Clay-level conjecture.
12.2. Track progress via formal benchmarks & peer review.
12.3. Human-gate any frontier theorem before release.

13. Phase 4: Generalist Mathematician (2031+)
13.1. ≥95 % FrontierMath; full IMO gold.
13.2. Publish referee-verified, AI-only theorems.
13.3. Enable on-demand axiom switching or proposal.

14. Safety & Alignment
14.1. Independent board for dual-use risk (cryptography, bio-math).
14.2. Log every search trace; generate explainable NL+diagrams.
14.3. Human-gate sensitive conjectures.

15. Scale Resources & Teams
15.1. Expand from 1–2 K to 10–30 K GPUs as self-play scales.
15.2. Grow formal corpora from <2 TB to 20–50 TB with autoformalized arXiv.
15.3. Recruit 100+ AI researchers, formalizers, and mathematicians.

16. Success Criteria
16.1. Continuous metrics: FrontierMath, miniF2F, ProofNet.
16.2. Breakthrough: AI contributes to open problems; AI-only published key results.
16.3. Safety alarm: dual-use thresholds require human oversight.

23.04.202512:55

Nvidia open-sourced Describe Anything! It can generate detailed descriptions for user-specified regions in images and videos, marked by points, boxes, scribbles, or masks.

Project page: https://describe-anything.github.io/

23.04.202512:51

Dia: SOTA open-weights TTS model for ultra-realistic dialogue.

Dia directly generates highly realistic dialogue from a transcript. You can condition the output on audio, enabling emotion and tone control. The model can also produce nonverbal communications like laughter, coughing, clearing throat, etc.

GitHub: https://github.com/nari-labs/dia/

23.04.202512:01

Stock market return in India 🇮🇳 vs in China 🇨🇳. From this graph, you'd think that India is the most promising emerging industrial power while China is an agricultural basket case.

23.04.202509:23

It would be unwise for Ukraine to agree to Trump's peace plan.

There are no security guarantees, and Trump refuses to pressure Russia while threatening Ukraine. In fact, he has refused to sell Ukraine air defense systems, even though Kyiv is prepared to cover the cost. Trump's pro-Russian stance goes as far as unwinding efforts to investigate Russian war crimes and forcing Kyiv to recognize territories that not even China recognizes as Russian.

In other words, Ukraine would not benefit from such a deal. In fact, Ukraine would be extorted for its resources by a government now clearly allied with its mortal enemy. Meanwhile, the tens of thousands of Ukrainian children who were forcibly transferred—or effectively kidnapped—to Russia will remain there and be brainwashed to hate their own country.

Such an arrangement would likely lead to several detrimental outcomes:

1. Russia would use the sanctions relief to rebuild its military capacity for a much bigger attack in the future.

2. Fearing this renewed attempt to destroy Ukraine, millions would flee. Because of this threat, necessary investments would not be made.

3. Russian interference in Ukrainian politics would undermine national cohesion.

Currently, Ukraine is in a much stronger position than it will be in a few years, when Russia has had time to prepare for a renewed assault.

This consideration extends to European security interests as well, as the conflict currently constrains Russian resources that might otherwise be directed toward other European nations. Every Russian soldier killed now is one less that Europeans will have to fight in the future.

22.04.202519:39

Physical Intelligence got a robot to clean up homes that were never seen in its training data! Their new model, π-0.5, aims to tackle open-world generalization.

Physical Intelligence took their robot into homes that were not in the training data and asked it to clean kitchens and bedrooms.

Read more: https://www.pi.website/blog/pi05

22.04.202519:33

The time horizons of coding agents are doubling every ~4 months.

Read more: https://theaidigest.org/time-horizons

22.04.202516:48

Links for 2025-04-22

AI

1. AI assisted search-based research actually works now https://simonwillison.net/2025/Apr/21/ai-assisted-search/

2. Pre-training isn't dead, it’s just resting https://tmychow.substack.com/p/pre-training-isnt-dead-its-just-resting

3. Questions about the Future of AI: Considerations about economics, history, training, deployment, investment, and more https://www.dwarkesh.com/p/questions-about-ai

4. Demis Hassabis discusses AGI, AI in medicine, and DeepMind’s groundbreaking advancements in robotics and reasoning. https://www.cbsnews.com/video/demis-hassabis-ai-deepmind-60-minutes-video-2025-04-20/

5. AgentA/B: Automated and Scalable Web A/BTesting with Interactive LLM Agents https://arxiv.org/abs/2504.09723

6. Let Me Grok for You: Accelerating Grokking via Embedding Transfer from a Weaker Model https://arxiv.org/abs/2504.13292

7. Eagle 2.5: Boosting Long-Context Post-Training for Frontier Vision-Language Models https://nvlabs.github.io/EAGLE/

8. A Deep Peek into DeepSeek AI’s Talent and Implications for US Innovation https://www.hoover.org/research/deep-peek-deepseek-ais-talent-and-implications-us-innovation

Brains

1. Each of the Brain’s Neurons Is Like Multiple Computers Running in Parallel https://singularityhub.com/2025/04/21/each-of-the-brains-neurons-is-like-multiple-computers-running-in-parallel/

2. “It looks like most mammals, at least most large mammals, have the brains they need, while primates, especially large primates, have the brains they can afford.” https://logarithmichistory.wordpress.com/2024/04/19/ground-up-monkey-brains-8/

3. Geoffrey Hinton says the more we understand how AI and the brain actually work, the less human thinking looks like logic. We're not reasoning machines, he says. We're analogy machines. We think by resonance, not deduction. https://youtu.be/vpUXI9wmKLc?si=E_t6BgVIhrdI3JN2&t=1703

Miscellaneous

1. “The semiconductor manufacturing business is absolutely immense. To give the numbers some perspective, in 2024, chip makers generated revenues that were about three quarters of the size of the US defense budget and about two-thirds the size of the social services budget allocated by Congress.” https://www.nextplatform.com/2025/04/21/the-chips-are-definitely-not-down/

2. A novel hybrid digital–analog approach to quantum simulation that we used to emulate quantum behavior of a magnet, resulting in a notable scientific discovery and new possibilities for beyond-classical applications. https://research.google/blog/a-new-hybrid-platform-for-quantum-simulation-of-magnetism/

3. Multiplex Gene Editing: Where Are We Now? https://www.lesswrong.com/posts/oSy5vHvwSfnjmC7Tf/multiplex-gene-editing-where-are-we-now

4. Last month, the Nucleic Acid Observatory identified an engineered virus sequence in wastewater. https://naobservatory.org/blog/updates-apr-2025/

5. KAIST Develops Retinal Therapy to Restore Lost Vision https://www.kaist.ac.kr/site/newsen/html/news/?mode=V&mng_no=45350

6. EthnoGuessr is a GeoGuessr variant: it shows you pictures of an ethnic group, you click on the map where you think they’re from. https://hbd.gg/

22.04.202515:11

Usually US economic pain is cushioned by falling bond yields and a strengthening dollar, which mean lower interest rates and more spending power for consumers.

This time we’re seeing the opposite, meaning the pain will be amplified.

Basically what normally happens is investors think “Stocks are too risky now, so let’s shift into US bonds and the dollar, which are a safer bet because America is a stable and well-run country with a good handle on its deficit and inflation.”

This time? Not so much.

The rise in treasury yields since Trump’s tariffs were announced leads to an increase in US debt interest payments that is larger than all the DOGE savings.

— John Burn-Murdoch

22.04.202514:43

Ukraine’s Biggest Hits: High‑Value Military & Energy Strikes (2022 – 2025)

24 Mar 2022 – Landing ship *Saratov* (Berdiansk): Tochka‑U missile sank the LST.
1 Apr 2022 – Belgorod fuel depot: Mi‑24 helicopters rocketed tanks, forcing fuel‑convoy reroutes.
14 Apr 2022 – Flagship *Moskva* (Black Sea): Neptune missiles set the cruiser ablaze; it capsized.
25 Apr 2022 – Bryansk pipeline terminal: Explosions halted Druzhba oil flows briefly.
22 Jun 2022 – Novoshakhtinsk refinery (Rostov): Drones destroyed a distillation unit; plant halted.
9 Aug 2022 – Saky air base (Crimea): Blasts destroyed 11 jets and ammo stocks.
8 Oct 2022 – Kerch Bridge (Crimea): Truck bomb collapsed two road spans and torched rail tankers.
29 Apr 2023 – Sevastopol fuel depot (Crimea): Sea drones torched 40 000 t of naval diesel.
4 Aug 2023 – *Olenegorsky Gornyak* landing ship (Novorossiysk): Sea drone holed the LST.
30 Aug 2023 – Pskov air base: Drone swarm destroyed two Il‑76 transports deep inside Russia.
13 Sep 2023 – Dry dock (Sevastopol): Storm Shadow missiles sank *Minsk* and damaged *Rostov‑on‑Don*.
22 Sep 2023 – Black Sea Fleet HQ (Sevastopol): Cruise missile blast struck the command center.
17 Oct 2023 – Berdiansk & Luhansk airfields: ATACMS cluster missiles destroyed nine helicopters.
26 Dec 2023 – Landing ship *Novocherkassk* (Feodosiya): Cruise missiles exploded the vessel.
14 Jan 2024 – A‑50 AWACS & Il‑22 (Sea of Azov): Patriot SAMs downed the AWACS and damaged the Il‑22.
21 Jan 2024 – Ust‑Luga export terminal: Long‑range drones halted Russia’s largest fuel hub.
25 Jan 2024 – Tuapse Rosneft refinery (Krasnodar): Drones ignited a vacuum column; plant idle for months.
14 Feb 2024 – Landing ship *Caesar Kunikov*: Sea drones sank the Ropucha‑class LST off Crimea.
23 Feb 2024 – Second A‑50 AWACS (Krasnodar): S‑200 SAM shot down another AWACS.
5 Mar 2024 – Patrol ship *Sergey Kotov* (Kerch Strait): Drone swarm sank the patrol vessel.
12 Mar 2024 – Kstovo (NORSI) refinery: 50‑drone raid wiped out half of the plant’s output.
13 Mar 2024 – Ryazan refinery: Strike knocked out 70 % of capacity.
17 Mar 2024 – Slavyansk‑on‑Kuban refinery: Drones ignited tanks; operations suspended.
28 Mar 2024 – Kuibyshev refinery (Samara): Strike halted output; prompted fuel imports from Belarus.
2 Apr 2024 – Taneco refinery (Tatarstan): Long‑range drone struck the Urals plant.
19 Apr 2024 – Tu‑22M3 bomber (Stavropol): S‑200 SAM downed the Backfire in flight.
5 Jun 2024 – Novoshakhtinsk refinery #2: Drone destroyed 1.5 Mt of oil/products and halted the plant.
24 Aug 2024 – Ostrogozhsk ammo depot (Voronezh): Drones detonated 5 000 t of munitions.
7 Sep 2024 – Soldatskoye depot (Voronezh): Drones destroyed a KN‑23 missile warehouse.
18 Sep 2024 – Toropets 107th GRAU arsenal (Tver): Drone strike caused a 2.8‑M quake by detonating 30 000 t.
21 Sep 2024 – Tikhoretsk ammo base (Krasnodar): Strike leveled 90 % of the 719th artillery depot.
9 Oct 2024 – Karachev 67th GRAU arsenal (Bryansk): Drones triggered massive secondary explosions.
19 Nov 2024 – ATACMS & drones, Karachev: Ballistic missiles and drones re‑hit the arsenal, causing a giant fireball.
29 Nov 2024 – S‑400 battery (Simferopol): Strike destroyed a launcher and 92N6 radar.
30 Dec 2024 – 76th VDV HQ (Lgov, Kursk): Storm Shadow killed senior airborne officers in HQ.
3 Feb 2025 – Volgograd refinery & Astrakhan gas plant: Drone swarm set both ablaze, suspending flights.
19–20 Mar 2025 – Engels‑2 ammo dump (Saratov): Drone swarm detonated 96 cruise missiles, devastating storage.
22 Apr 2025 – 51st GRAU Arsenal ammo depot (Oryol Oblast): Ukraine struck one of Russia’s largest munitions depots, setting off massive secondary explosions that continue to rock the facility.

22.04.202509:19

21.04.202517:29

Links for 2025-04-21

AI

1. A new agentic approach called 'streams' will let AI models learn from the experience of the environment without human 'pre-judgment'. DeepMind scholars David Silver and Richard Sutton published the most important essay on AI since The Bitter Lesson: "Welcome to the Era of Experience". [PDF] https://storage.googleapis.com/deepmind-media/Era-of-Experience%20/The%20Era%20of%20Experience%20Paper.pdf

2. OpenAI's o3: Over-optimization is back and weirder than ever https://www.interconnects.ai/p/openais-o3-over-optimization-is-back

3. "Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?" https://natolambert.substack.com/p/does-reinforcement-learning-really

4. LLM responses to benchmark questions are getting longer over time https://epoch.ai/data-insights/output-length

5. “Why Should I Assume CCP AGI is Worse Than USG AGI?” https://www.lesswrong.com/posts/MKS4tJqLWmRXgXzgY/why-should-i-assume-ccp-agi-is-worse-than-usg-agi-1

6. Skyrocketing AI Intelligence: ChatGPT's o3 sets new record https://www.maximumtruth.org/p/skyrocketing-ai-intelligence-chatgpts

7. AI 2027 is a Bet Against Amdahl's Law https://www.lesswrong.com/posts/bfHDoWLnBH9xR3YAK/ai-2027-is-a-bet-against-amdahl-s-law

8. “Things will go slow until they go fast. Really fast. Things feel fast today, but I think we are actually still accelerating and we will actually start to go even faster. Buckle up.” https://x.com/TheRealAdamG/status/1913998366632968381

9. Is Gemini now better than Claude at Pokémon? https://www.lesswrong.com/posts/7mqp8uRnnPdbBzJZE/is-gemini-now-better-than-claude-at-pokemon

10. Research Notes: Running Claude 3.7, Gemini 2.5 Pro, and o3 on Pokémon Red https://www.lesswrong.com/posts/8aPyKyRrMAQatFSnG/untitled-draft-x7cc

11. Making AI-generated code more accurate in any language https://news.mit.edu/2025/making-ai-generated-code-more-accurate-0418

12. Microsoft researchers say they’ve developed a hyper-efficient AI model that can run on CPUs https://techcrunch.com/2025/04/16/microsoft-researchers-say-theyve-developed-a-hyper-efficient-ai-model-that-can-run-on-cpus/

13. No NVIDIA? No problem! Huawei trains a strong dense model on Ascend NPUs https://arxiv.org/abs/2504.07866

14. New AI Startup Aims to Replace All Human Workers https://www.mechanize.work/

Miscellaneous

1. An international team led by Fabian Garmroudi has succeeded in producing new, efficient thermoelectric materials that could compete with state-of-the-art materials, offering greater stability and lower cost. https://www.tuwien.at/en/all-news/news/neue-hybridmaterialien-als-effiziente-thermoelektrika

Усе пасты

Рэкламаваўся0 секунда

23.04.202509:49Catacombs of Dissent

Рэкламаваўся0 секунда

20.04.202514:49Wiggerlab

112

Рэкламаваўся0 секунда

18.04.202518:13/slashbiz/ ~ Sneed, Crypto, Business and Finance, AI, memes and news, and being poor in general

3.7K

Рэкламаваўся0 секунда

18.04.202518:08/biz/ Crypto MayMay Shithole - Bogdanof Down! I repeat both Bogdanof down!

2.9K

Рэкламаваўся0 секунда

18.04.202512:41Wiggerlab

112

Рэкламаваўся1 дзень

17.04.202519:55Wiggerlab

112

Рэкламаваўся0 секунда

17.04.202502:07Machinae Animatae #УкрТґ

102

Рэкламаваўся0 секунда

10.04.202509:31Wiggerlab

112

Рэкламаваўся0 секунда

07.04.202507:20CRITICAL//Consensus

1.9K

Падрабязней

Рэкорды

15.04.202523:59

3.8KПадпісчыкаў

16.04.202518:33

6Індэкс цытавання

24.11.202423:59

1.7KАхоп 1 паста

24.11.202423:59

2.9KАхоп рэкламнага паста

17.04.202519:56

38.10%ER

04.04.202523:59

24.60%ERR

Развіццё

Падрабязней

Падпісчыкаў

Індэкс цытавання

Ахоп 1 паста

Ахоп рэкламнага паста

ERR

Падрабязней

Папулярныя публікацыі Axis of Ordinary

Усе пасты

19.04.202519:07

A behind the scenes look at an Ukrainian Drone Factory owned by a company called SkyFall.

They make over 4,000 drones per day with a drone being completed every 27 seconds.

10.04.202509:01

Thanks to Google, we now have clearly superhuman AI at medical diagnosis.

It’s so good that clinicians do *worse* when assisted by the AI compared to if they just let the AI do its job.

Articulate Medical Intelligence Explorer: A research AI system optimized for diagnostic reasoning and conversations

Paper 1: https://goo.gle/4lpQ8xg
Paper 2: https://goo.gle/3G4DNPe

In the first paper, AMIE matched or exceeded primary care physicians in a study of simulated text-chat medical consultations with validated patient actors.

The second paper evaluated AMIE’s ability to generate a differential diagnosis alone or as a clinician aide. The authors looked at 303 challenging, real-world medical cases. AMIE outperformed clinicians, and clinicians assisted by AMIE showed improved performance.

15.04.202519:15

A friendly reminder that in 2022 published AI researchers thought AI wouldn't be able to "write simple python code" until 2027. Or remember when AI couldn't draw hands? That was 2023. People are still underestimating AI-progress.

https://x.com/Miles_Brundage/status/1911878309329457207

23.04.202516:34

"You Don't Have the Cards."

17.04.202508:12

And that is nothing. We'll probably see the release of o5 and possibly even o6 this year. And that's just the beginning.

A lot of people are still deluding themselves because they're scared shitless about this development. Just look at the retards in the comments here. But reality doesn't disappear when you close your eyes.

14.04.202512:45

No joke.

15.04.202511:18

Google DeepMind, David Silver reveals:

We built a system that used RL to discover its own RL algorithms.

This AI-designed system outperformed all human-created RL algorithms developed over the years.

P.S. Note that Google has been using AI for years to design its computer chips (TPUs), which in turn are used to train better AI models. AlphaChip accelerates and optimises chip design and creates superhuman chip layouts.

Original source for the video: https://www.youtube.com/watch?v=zzXyPGEtseI

18.04.202511:42

“The Russian Paradox: So Much Education, So Little Human Capital”

Read more: https://conversableeconomist.com/2025/04/17/russia-an-unhealthy-population/

17.04.202521:23

o3 crops and zooms into the image.

Finds every little clue and runs quick searches.

Then drops the exact map coordinates.

08.04.202516:50

In the long term, there's really only one way to prevent the global economy from slipping into a sustained recession: continued progress in AI and robotics. The technology that could kill us all. The only question seems to be whether humanity will go out with a bang or a whimper.

13.04.202511:58

Gmail Creator Paul Buchheit: We're On The Path Towards AGI

The key point that happened is that we crossed a line where AI went from a research project, where you put in a lot of money and don't get much out, to a thing where you put in money and get out more.

It's like when a reaction goes critical ... that's where we are right now: people can't seem to throw money at it fast enough.

Source: https://www.youtube.com/watch?v=LSUviaN1eso

26.03.202522:37

It's pretty amazing that Grok is consistently pointing out how Musk is spreading lies. Gives me an iota of hope that friendly AI might just be the default outcome.

https://x.com/grok/status/1904798600409853957

16.04.202515:27

Why Human Intelligence May Be Just a Glimmer of What's Possible

The possibility of vastly superhuman Artificial Intelligence (AI) is often seen as extraordinary, but perhaps it should be our default assumption. Comparing the constraints on human intelligence from its evolutionary origin with AI's potential suggests a significantly higher ceiling for artificial cognition.

Human intelligence arose via natural selection – a blind, incremental process optimizing only for immediate survival and reproduction, not maximal intelligence. Biological constraints like metabolic costs and childbirth risks limited brain development. Evolution is path-dependent; superior cognitive designs might be unreachable through gradual mutation from our biological base. The lack of species clustered just below human intelligence also argues against a nearby natural ceiling.

AI, especially future AGI, transcends these limits:

- Speed: Biological neurons (~10^2 Hz) are vastly slower than digital processors (10^9 Hz) or future optical computing. AI communication speeds (10^11 bps) dwarf human language (~10-50 bps), enabling faster thought and collaboration.

- Scalability & Duplicability: A capable AI can be copied almost instantly, unlike the decades needed to train human experts. This allows massive scaling of "talent" by turning capital into compute power, replicating individuals or entire expert teams on demand.

- Editability & Self-Improvement: AI can have 'root access' for modification. Systematic debugging, experimentation on copies, and version control enable rapid recursive self-improvement impossible for biological "wetware." AI can iterate and improve its own architecture deliberately.

- Memory: AI can possess vastly larger working memory than humans (e.g., holding libraries vs. a phone number) and near-perfect, instantly searchable long-term recall, free from human biases and decay.

- Collective Intelligence: AIs can merge knowledge directly and communicate with perfect fidelity, creating "collective brains" that learn and innovate far faster than human groups reliant on slow, lossy language. Insights can propagate instantly across countless copies.

- Architectural Freedom: AI isn't restricted to evolved brain structures; designers can explore fundamentally different and potentially superior cognitive architectures optimized for various goals.

- Focus & Endurance: Lacking biological needs (like sleep) and inherent cognitive biases, AI can operate 24/7 with unwavering focus and potentially greater rationality and consistency towards its objectives.

Human intelligence, though remarkable, is limited by its contingent evolutionary past. AI's fundamental advantages in speed, scalability, editability, memory, collective function, and architectural freedom strongly suggest its potential cognitive ceiling is vastly higher. Superhuman general AI thus appears a plausible, even likely, outcome of continued technological progress, rather than mere speculation.

03.04.202519:34

How, exactly, could AI take over by 2027? by Daniel Kokotajlo and Scott Alexander

We predict that the impact of superhuman AI over the next decade will be enormous, exceeding that of the Industrial Revolution.

We wrote a scenario that represents our best guess about what that might look like. It’s informed by trend extrapolations, wargames, expert feedback, experience at OpenAI, and previous forecasting successes.

The scenario: http://ai-2027.com
PDF version: https://ai-2027.com/scenario.pdf
Podcast: https://www.dwarkesh.com/p/scott-daniel
YouTube: https://www.youtube.com/watch?v=htOvH12T7mU

16.04.202518:13

OpenAI releases o3 and o4-mini and they feature new emergent capabilities.

We didn't train the model to use certain strategies directly. We didn't say 'simplify your solution' or 'double check'. It just organically learns to do these things.

Highlights:

- The models can agentically use and combine every tool within ChatGPT—this includes searching the web, analyzing uploaded files and other data with Python, reasoning deeply about visual inputs, and even generating images.

- In evaluations by external experts, o3 makes 20 percent fewer major errors than OpenAI o1 on difficult, real-world tasks.

- At equal latency and cost with OpenAI o1, o3 delivers higher performance.

- For the first time, these models can integrate images directly into their chain of thought. They don’t just see an image—they think with it.

Read more: https://openai.com/index/introducing-o3-and-o4-mini/

Увайдзіце, каб разблакаваць больш функцый.