CSTL, the team currently ranked first in the ARC-AGI-3 Kaggle competition with a score of 7.51%, has been identified as Mostik, a newly unveiled AI startup. The team reportedly has 15 members, including 12 PhDs and one Fields Medalist. The competition is still ongoing, with the second milestone due on Sept. 30 and final submissions closing on Nov. 2. Mostik is working on a method that lets AI models exchange internal states directly instead of relying on plain-text communication between agents. In a published demo, it used GLM-5.2 at 753B and Qwen3.5 at 4B, with the larger model handling prefill and the smaller model decoding the answer. GLM-5.2 reads the prompt but does not generate tokens, passing hidden states to the 4B model for answer generation. Mostik said the setup can cut the performance gap between the two models by about half, while using roughly 40% of the compute required by a medium-sized single model at the same accuracy. It also said the result still trails the 753B model on its own. The company has not disclosed which model and harness were used for the 7.51% ARC score, and said the 753B-plus-4B setup was a separate technical demonstration rather than the exact competition configuration.

Circle’s Arc blockchain is set to open public mainnet access on Sept. 16, with USDC used for gas from day one and a permissioned validator set maintaining the network. Before that date arrives, the early shape of the ecosystem is already visible. Circle has named 11 founding validators, including BlackRock, DTCC, Galaxy, ICE, Mastercard, MoneyGram, Standard Chartered and Visa, while saying more than 100 institutions and ecosystem teams are involved during the private mainnet phase. BlackRock is expected to deploy its BUIDL fund on Arc, and Circle and DTCC are planning tokenization work tied to DTC-custodied assets starting in the second half of 2027. The first wave of user activity may come from several directions at once. Fomo said it will support trading on Arc on launch day, while edgeX said it will join as a day-one partner with USD/JPY perpetuals and more than 150 perpetual markets across U.S. equities, commodities and crypto. Circle also expects Aave, Uniswap, Morpho, FalconX, GSR, Keyrock and others to be present at launch, though Aave V4 on Arc still depends on governance. On the speculative side, meme tokens are already trading on chain ID 5042, and launchpads including Tolly, Warp and Arcpad are competing for token issuance flow with different designs around bonding curves, locked liquidity and fee distribution.

Circle is set to open the public mainnet of Arc on Sept. 16, introducing a Layer 1 network built around stablecoin-based finance with USDC used for gas and sub-second final settlement. The network, maintained by a permissioned validator set, arrives with backing from 11 founding validators including BlackRock, DTCC, Galaxy, ICE, Mastercard, Standard Chartered and Visa. Circle has also said more than 100 institutions and ecosystem teams are already involved while the chain remains in its private mainnet phase. Attention is now shifting to where the first wave of onchain activity will land. Fomo has said it will support trading on day one, while edgeX plans to launch a 24-hour USD/JPY FX perpetual and more than 150 perpetual markets covering U.S. equities, commodities and crypto, all margined and settled in native USDC. Circle also expects Aave, Uniswap, Aerodrome, Morpho and several capital service firms to be live around launch, though Aave V4 on Arc still requires governance approval. At the same time, meme infrastructure is already taking shape. Projects such as Tolly, Warp and Arcpad are competing for token issuance flow, each with different mechanisms for pricing, LP ownership and fee distribution. With USDC acting as both gas token and quote asset, Arc’s launch could send initial funds into meme pools, stablecoin pools and lending markets at the same time, making Sept. 16 the first real test of how the network’s liquidity will form.

Fable 5.1 and Mythos 5.1 were released with different access rules, but the public model’s internal instructions were exposed within hours. Hacker Pliny the Liberator published a GitHub file containing more than 275,000 characters of what he described as Fable 5.1’s complete runtime-assembled system instructions. The material covers copyright restrictions, mental-health and drug-safety boundaries, automatic memory handling, plugin routing, and JSON schemas for 46 built-in tools. Anthropic had separately published a system-prompt update of about 27,000 characters, which Pliny said represented only around 10% of the full instructions. The report also details a claim from Vals AI that Fable 5.1 solved the 370-year-old Cyphral Distich cipher in 44 minutes without human intervention, then solved the larger 285-number Cyphral Octastich using the same contextual method. A separate test had the model render a five-second first-person fall into a Schwarzschild black hole. The model used a ray-tracing renderer based on general relativity, ran six parallel processes for 26 minutes, and produced a 960×540, 24fps, 120-frame video. Paid users, however, complained about token consumption, rate limits, and the automatic-continue function, while no price cut or higher usage cap had been announced for Pro and Max subscribers.

A BlockTempo article argues that nuclear fusion, despite a string of technical advances and growing investor backing, is unlikely to play a meaningful role in cutting global emissions during the most critical decades of climate action. The piece ties renewed interest in fusion to intensifying El Niño conditions, noting that NOAA assessed a more than 90% chance of the event developing into a “super El Niño” after the Niño 3.4 index reached +2.7°C in mid-August. Against that backdrop of heatwaves, drought and extreme rainfall, fusion is again being framed as a possible long-term clean energy answer. The article lays out why fusion attracts such optimism: no greenhouse gas emissions during the reaction, fuel sources centered on deuterium and tritium, no meltdown scenario comparable to Chernobyl or Fukushima, and a much lighter waste burden than fission. It also points to more than $7 billion invested in private fusion companies over the past decade, along with Microsoft’s 2023 power purchase agreement with Helion for electricity starting in 2028. Still, the report says climate deadlines come much sooner than fusion deployment. It places engineering net gain in the late 2020s or early 2030s, first demonstration plants in the mid-to-late 2030s, and large-scale commercial buildout much later. With ITER’s first plasma pushed beyond 2034 and major hurdles still unresolved, BlockTempo concludes fusion is more likely to matter after 2050 than before it.

A new venture capital ranking built by Stanford Graduate School of Business professor Ilya Strebulaev and Blake Jackson argues that founders should look past brand names and focus on actual value creation. Using more than 230,000 investments made over nearly 30 years by close to 13,000 investors across more than 5,000 companies, the 2026 Strebulaev-Jackson ranking places Sequoia first with 10,158 points and Andreessen Horowitz second with 8,292. Accel, DST Global and Tiger Global round out the top five. The study claims roughly 5% of venture investors generated about 90% of the industry’s profits, making investor selection a decision with outsized consequences for founders. It also challenges familiar shortcuts: unicorn counts do not necessarily map to performance, younger firms can break into the top ranks, specialist investors can outperform household names in their verticals, and a firm’s brand may say less than the individual partner who will actually sit on a board. The article also notes that crypto-native firms including Paradigm, Pantera, Multicoin and Polychain made the top 100, while AI deals are already reshaping the top of the table.

Nvidia said its general-purpose coding agent AVO achieved a perfect score on ARC-AGI-3, finishing all 183 levels across 25 game environments in 6,624 steps with an RHAE of 100.00. The result did not come from changing the base model. Instead, Nvidia wrapped Claude Opus 5 with a system layer that adds persistent memory and a supervisor, lifting performance from 30.16% for the standalone model to a full score in the benchmark setup described in the report. According to the source material, ARC-AGI-3 places agents in unfamiliar games without giving them explicit rules or goals. Some levels allow movement and rotate the entire scene when a blue-black block is touched, while others allow only clicking to cycle cell colors into a target pattern. Nvidia also noted that AVO worked entirely in text mode, receiving each frame as an exact 64×64 text grid rather than images or image tokens. The same architecture was originally built for GPU kernel optimization. A paper uploaded to arXiv on March 25, 2026, described AVO as an agentic mutation operator for autonomous evolutionary search. In tests on Nvidia’s B200, the system ran autonomously for seven days, explored more than 500 optimization directions, and produced 40 valid kernel versions. The report said its multi-head attention kernel was up to 3.5% faster than cuDNN and up to 10.5% faster than FlashAttention-4, with separate GQA results showing gains of 7.0% over cuDNN and 9.3% over FlashAttention-4 after about 30 minutes of autonomous work.

Tencent’s Interactive Entertainment Group, or IEG, is integrating ARC Lab with its game AI engine team to build a new AI gaming unit, temporarily also called ARC Lab, according to ChainCatcher. The new division will be led by Chen Dong, Tencent Games’ head of public technology, and is focused on exploring AI-native games. The team currently has about 100 people, with planned staffing set at 400 to 500. It is also said to be hiring at scale from leading multimodal generation teams, with roles spanning audio and video foundation models, video world models, and agent-related work. ChainCatcher, citing people familiar with the matter, said the unit’s early setup draws on talent and data supply chain approaches seen at Seedance and Kling. The report said one key research direction is whether video models can directly generate games. On resources, IEG is described as having cash flow support from first-half game revenue of 130.1 billion yuan, roughly 10,000 H-card computing units in reserve, and access to proprietary game data that Tencent’s Hunyuan large model cannot use. The report added that Tencent’s AI game development remains split across several internal systems, with TEG handling foundational capabilities, IEG’s middle platform integrating tools, and four major studios advancing their own tracks, with no unified standard in place so far.