#EpochAI
LightNews LightNews
wtgowers.bsky.social
It seems that OpenAI's latest model, o3, can solve 25% of problems on a database called FrontierMath, created by EpochAI, where previous LLMs could only solve 2%. On Twitter I am quoted as saying, "Getting even one question right would be well beyond what we can do now, let alone saturating them."
December 20, 2024 at 11:15 PM
pip.net
Marked the fraction of DC cost that data centers in Space _might_ reduce while significantly complicating the remaining 90% of line items. Convincing proposition.

[Data: EpochAI]
June 26, 2026 at 6:12 AM
christianstump.bsky.social
New math benchmark from math.science-bench.ai :

209 research-level mathematics problems from Combinatorics, Algebra, Geometry, Number Theory, and others.

👉 math.science-bench.ai/benchmarks/

#AI #Mathematics #AIBenchmark #EpochAI #FrontierMath #OpenAI #Gemini #Grok
November 1, 2025 at 2:18 PM
airrationality.bsky.social
EpochAI is next to throw in the gauntlet in the "solving open problems with AI" game
open.substack.com/pub/aichats/...
EpochAI is next to throw in the gauntlet in the "solving open problems with AI" game
A critical review of the recent claims by EpochAI that frontier models solved an open FrontierMath problem with the assistance of Gemini.
open.substack.com
March 25, 2026 at 9:48 AM
diack.bsky.social
Epochai unveils FrontierMath, an AI benchmark for advanced mathematical reasoning.
epochai.org/frontiermath
FrontierMath
epochai.org
November 8, 2024 at 10:46 PM
stephenturner.us
EpochAI releases a database of over 1,100 biological AI models across nine categories, analyzing their safeguards, accessibility, training data sources, and the foundation models they build on epoch.ai/blog/expandi... #biosecurity
March 6, 2026 at 7:00 PM
ramon-astudillo.bsky.social
* = They "FrontierMath/EpochAI"
January 19, 2025 at 8:23 PM
piotrsankowski.bsky.social
Model o3 rozwiązał trochę ponad 25% problemów ze zbioru EpochAI Frontier Math (epoch.ai/frontiermath), gdzie każdy problem wymaga godzin pracy od ekspertów matematyków. Niedawno mówiłem, że matematyka będzie pierwszą dziedziną, gdzie AI będzie wspierać badania naukowe (www.pap.pl/aktualnosci/...)
December 20, 2024 at 8:57 PM
theo-mir-rt.selfhosted.social
RT @AuthoStats: AA Intelligence Index uses their own harness. Most costs are from 3 Benchmarks, and the bulk of Groks trouble is Terminal Bench 4, which uses its own miniSWE harness and is flagged as a flawed benchmark by EpochAI.
September 22, 2026 at 5:31 AM
biomattrs.bsky.social
Example problem from EpochAI FrontierMath.
epoch.ai/frontiermath...

From Wikipedia ...
"The conjecture was made by Emil Artin to Helmut Hasse on September 27, 1927, according to the latter's diary. The conjecture is still unresolved as of 2024."

We're gonna need a new frontier 🥹
December 20, 2024 at 11:55 PM
gingiu.bsky.social
Molto interessante
#openai #datacenter #epochai
open.substack.com/pub/artifici...
AI data centers - What you need to know #59
Series “Laboratory”
open.substack.com
November 15, 2025 at 1:25 PM
arnaudvineski.bsky.social
Si Frandroid et EpochAI le confirment dans une étude financée par eux-mêmes alors, c’est que ça doit être vrai …
December 12, 2025 at 10:40 AM
sbhasin.bsky.social
India is racing to become an AI superpower, but at what cost to our grid?

#AI #EnergyCrisis #India #Tech #Sustainability #DigitalIndia #MintDataBites #EpochAI
July 22, 2026 at 4:53 AM
ramon-astudillo.bsky.social
☝️EpochAI person also does not state that the test set they showed numbers in was the mentioned blind set ... but this could be just a misshap (hopefully)
January 19, 2025 at 7:47 PM
notengoprisa.bsky.social
Era de EpochAI si no recuerdo mal. 7 millones sería la capacidad de computación que tienen ahora mismo, que es muy poca comparado a lo que tienen anunciado y en construcción.
October 14, 2025 at 2:09 PM
epochai.bsky.social
Transcript available here: open.substack.com/pub/epochai...
Democratic proofs or fake proofs? - Jesús De Loera
A proof only 15 experts understand is less valuable than one any undergraduate can verify using a computer.
epochai.substack.com
October 13, 2025 at 8:50 PM
handackett.bsky.social
...which would mean it should be pretty easy to estimate utilization at present, because the accounting only needs to cover big centralized blobs of compute

I wonder if the EpochAI people have looked into this
August 19, 2026 at 12:22 AM
ketanjoshi.co
Took 42 GWh to train GPT4 - IEA estimate ~1.7 terawatt hours to train all the large frontier models from 2020 to now.

Like with so many of the numbers in this report it's pretty big, but small on top of the denominator of 'everything'.
Model training Model training refers to the process of the model learning from data to identify relationships and patterns. Given a set of inputs, the model learns to generate an output, prediction or action that aligns with the patterns identified in the training data and with the model’s overarching objectives. In the last few years, the amount of data and calculations required to train state-of-the-art AI models has grown exponentially. For example, estimates put the training data for GPT-4 at around 4.9 trillion data points, and the training compute at around 22 trillion trillion calculations (that is, 2.2e25). Training is a time-consuming and energy-intensive process. Training calculations are performed on specialised computer chips such as GPUs. A single GPU can have a maximum rated power consumption of 1 000 watts in the case of the latest and most powerful chip. This is about as much as the power draw of a toaster. Large, state-of-the-art models are trained on clusters of many GPUs. For example, GPT-4 was trained on 25 000 GPUs with a combined rated power of around 10 MW (EpochAI, 2024). Additional power demand comes from information technology (IT) equipment operating alongside the GPUs in the servers used to train these models, such as CPUs, memory, networking equipment and switches.2 Adding the power demand of additional IT equipment and the cooling equipment gives a total rated power of the equipment used to train GPT-4 of around 22 MW. This is equivalent to the power draw of around 150 high-power electric vehicle charging stations. It is estimated that GPT-4 was trained for around 14 weeks. Taking a load factor of 84% (Shehabi, et al., 2024), this results in a training energy demand of around 42.4 gigawatt hours (GWh), or around 0.43 GWh per day of training. This is equivalent to the daily electricity consumption of around 28 500 households in advanced economies, or 70 500 households in emerging market and developing economies. After training, models ma… In the absence of better data, this gives us at least an order of magnitude for the total training electricity consumption of the large frontier AI models in the EpochAI dataset, which comes to around 1 700 GWh (1.7 TWh). This is equivalent to around 0.001% of global electricity consumption during this period from all sources, or 0.1% of the global electricity consumption of data centres during these years. Figure 1.15 ⊳ Estimated training-related maximum power draw, electricity consumption and cumulative electricity consumption for a set of large AI models since 2020
April 29, 2025 at 8:29 PM
arnaudvineski.bsky.social
Ils relaient une « étude » financée par EpochAI. Mais il est vrai que dans ce secteur, on confond souvent étude, campagne de com et étude de marché. Et que l’environnement on s’en fout comme d’une guigne
December 12, 2025 at 10:42 AM
alexvonwitzleben.bsky.social
#StargateofChina plan emerges to challenge US as #AI superpower in response to the strong lead held by the US over access to AI computing capacity. #EpochAI is estimating America has around three-quarters of the global computing power, compared to 15% in China on.ft.com/4mnJvLd via @FT #vonderLeyen
‘Stargate of China’ plan emerges to challenge US as AI superpower
Beijing increases oversight of its data centres to make better use of limited computing resources
on.ft.com
September 21, 2025 at 6:46 AM
peterwildeford.bsky.social
China's top AI models are trained on American chips. Preventing China's access to top American chips prevents their access to top models.

Using EpochAI data, we found only two examples of notable models trained using Chinese chips.
June 12, 2025 at 7:11 PM
bradluen.bsky.social
Reading the IEA's Energy and AI report

Didn't realize the energy that goes into training a mega model—they estimate the costliest model (Grok?) took 310 GWh. OTOH all models in the dataset combined cost 1.7 TWh, or 0.1% of global data center consumption over the period.
www.iea.org/reports/ener...
Figure 1.15 from IEA, Energy and AI (p. 42) Maximum training power draw of an AI model: 154 MW Maximum training electricity consumption of an AI model: 310 GWh Total training electricity consumption of 283 large AI models in the EpochAI dataset: 1700 GWh. "This is equivalent to around 0.001% of global electricity consumption during this period from all sources, or 0.1% of the global electricity consumption of data centres during these years."
September 9, 2025 at 5:48 AM
corrierenerd.bsky.social
Se ti sei mai imbattuto in modelli di intelligenza artificiale (IA) che sembrano risolvere problemi matematici complessi o rispondere a domande#ai #Anthropic #applicazioniIA #benchmarkIA #confrontomodelliIA #epoch #epochAI #google #GPQADiamond #IA #IAavanzata
www.corrierenerd.it/epoch-ai/
December 3, 2024 at 7:17 AM