Hasan Can retweetledi
Hasan Can
4.6K posts
Hasan Can
@HCSolakoglu
SWE & AI- News, Insights Posts in ENG&TUR Exploring AI
Proxima C B Katılım Temmuz 2020
2.5K Takip Edilen1.4K Takipçiler
At this price-performance level, M2.7 makes everything else look overpriced, including GPT-5.4 mini and Gemini 3 Flash. If benchmarks hold up in real world use, people will change their preferences. No one wants to waste their money.
Artificial Analysis@ArtificialAnlys
MiniMax has released MiniMax-M2.7, delivering GLM-5-level intelligence for less than one third of the cost MiniMax-M2.7 from @MiniMax_AI scores 50 on the Artificial Analysis Intelligence Index, an 8-point improvement over MiniMax-M2.5, which was released one month ago. This is driven by stronger performance on real-world agentic tasks and reduced hallucinations. MiniMax-M2.7 is now ahead of MiMo-V2-Pro (Reasoning, 49) and Kimi K2.5 (Reasoning, 47), and equivalent to GLM-5 (Reasoning, 50) while using 20% fewer output tokens and costing less than a third as much to run. MiniMax-M2.7 is a reasoning-only model and maintains the same per-token pricing as MiniMax-M2.5. Key takeaways: ➤ Strong performance on real-world agentic tasks: MiniMax-M2.7 achieves a GDPval-AA Elo of 1494, a significant improvement from MiniMax-M2.5 (1203) and ahead of MiMo-V2-Pro (Reasoning, 1426), GLM-5 (Reasoning, 1406), and Kimi K2.5 (Reasoning, 1283). It remains behind frontier models such as GPT-5.4 (xhigh, 1667) and Claude Opus 4.6 (Adaptive Reasoning, max effort, 1606) ➤ Reduced hallucinations: MiniMax-M2.7 scores +1 on the AA-Omniscience Index, up from MiniMax-M2.5 (-40). This is competitive with GPT-5.2 (xhigh, -1) and GLM-5 (Reasoning, +2), and well ahead of Kimi K2.5 (Reasoning, -8). The improvement from M2.5 is purely driven by reduced hallucinations, meaning the model is more likely to abstain from answering when it doesn’t know the answer, rather than guessing. M2.7 achieves a hallucination rate of 34%, lower than Claude Sonnet 4.6 (Adaptive Reasoning, max effort, 46%) and Gemini 3.1 Pro Preview (50%). ➤ Gains across most evaluations compared to MiniMax-M2.5: Outside of the GDPval-AA and AA-Omniscience improvements noted above, MiniMax-M2.7 improves in HLE (+9 p.p.), TerminalBench Hard (+5 p.p.), SciCode (+4 p.p.), IFBench (+4 p.p.), GPQA (+3 p.p.), and LCR (+3 p.p.). We saw a notable regression in τ²-Bench (-11 p.p.). ➤ Increased token use: MiniMax-M2.7 used ~87M output tokens to run the Artificial Analysis Intelligence Index, up 55% from MiniMax-M2.5 (~56M). It remains more token-efficient than other models such as GLM-5 (Reasoning, 110M) and Kimi K2.5 (Reasoning, ~89M) ➤ Leading cost efficiency: MiniMax-M2.7 cost $176 to run the Artificial Analysis Intelligence Index, maintaining the same $0.30/$1.20 per 1M input/output pricing as M2.5. This places it on the Pareto frontier of our Intelligence vs. Cost chart. For context, GLM-5 (Reasoning) cost $547 at equivalent intelligence, Kimi K2.5 (Reasoning) cost $371, and Gemini 3 Flash Preview (Reasoning) cost $278 Key model details: ➤ Context window: 200K tokens (equivalent to MiniMax-M2.5). ➤ Pricing: $0.30/$1.20 per 1M input/output tokens (unchanged from MiniMax-M2.5). ➤ Availability: MiniMax first-party API only. ➤ Modality: Text input and output only (no multimodality). ➤ Licensing: MiniMax has not announced whether MiniMax-M2.7 will be open weights. MiniMax-M2.5 is available under the MIT license.
English
🤖 Your favorite AI agents + Colab’s powerful cloud compute. 🤝
Colab's new Open Source MCP Server gives your agents programmatic access to control the entire notebook development lifecycle. Build, run, and visualize—securely in the cloud.
Dive into the details and the GitHub repo below 👇
Google for Developers@googledevs
Bringing Google Colab’s secure environment to your AI agents. ☁️🛠️ With the new open-source Colab MCP Server, your agents can now natively write and execute code inside a Colab Notebook. Secure, automated, and ready to go.
English
Hasan Can retweetledi
Hasan Can retweetledi
The newest model in the Mamba series is finally here 🐍
Hybrid models have become increasingly popular, raising the importance of designing the next generation of linear models.
We've introduced several SSM-centric ideas to significantly increase Mamba-2's modeling capabilities without compromising on speed. The resulting Mamba-3 model has noticeable performance gains over the most popular previous linear models (such as Mamba-2 and Gated DeltaNet) at all sizes.
This is the first Mamba that was student led: all credit to @aakash_lahoti @kevinyli_ @_berlinchen @caitWW9, and of course @tri_dao!
English
Hasan Can retweetledi
GPT-5.4 mini is available today in ChatGPT, Codex, and the API.
Optimized for coding, computer use, multimodal understanding, and subagents. And it’s 2x faster than GPT-5 mini.
openai.com/index/introduc…
English
@matvelloso It’s great, but the price increase really sucks. It cuts out a lot of the use cases of 2.5 flash lite.
English
gemini-3.1-flash-lite-preview is extremely underrated. I know I keep saying that, but nothing beats the (price*latency)/intelligence you get here.
English
Hasan Can retweetledi
Introducing 𝑨𝒕𝒕𝒆𝒏𝒕𝒊𝒐𝒏 𝑹𝒆𝒔𝒊𝒅𝒖𝒂𝒍𝒔: Rethinking depth-wise aggregation.
Residual connections have long relied on fixed, uniform accumulation. Inspired by the duality of time and depth, we introduce Attention Residuals, replacing standard depth-wise recurrence with learned, input-dependent attention over preceding layers.
🔹 Enables networks to selectively retrieve past representations, naturally mitigating dilution and hidden-state growth.
🔹 Introduces Block AttnRes, partitioning layers into compressed blocks to make cross-layer attention practical at scale.
🔹 Serves as an efficient drop-in replacement, demonstrating a 1.25x compute advantage with negligible (<2%) inference latency overhead.
🔹 Validated on the Kimi Linear architecture (48B total, 3B activated parameters), delivering consistent downstream performance gains.
🔗Full report:
github.com/MoonshotAI/Att…
English
linus was the original vibe coder before it was cool. dude just posts an angry email on the mailing list describing what he wants and thousands of engineers worldwide implement it for free. zero tokens consumed, zero API costs, infinite context window (30+ years of kernel knowledge). openai and anthropic are basically trying to replicate what linus has been doing with human contributors since 1991 except linus's agents dont hallucinate and they work for free @sahill_og
Sahil@sahill_og
Linus Torvalds created Linux at 21 without Claude or any other AI. - He didn't have a co-founder. - No VC funding. No office. - No team. - Just a personal project he posted to a mailing list: "I'm doing a free OS." 33 years later, it runs 97% of the world's servers, all smartphones, and the International Space Station. The most important software in history started as someone's side project. Absolute legend.
English
@bardozVAL I just tried it and realized I’m more consistent with this. Thanks for your contribution.
English
@HCSolakoglu you should make sure you account for how much extra the browser, peripherals, monitor, pc, and human benchmark's specific implementation add. github.com/sirbardo/laten… I always get way less on this much faster implementation I made compared to human benchmark.
English
Fastest possible human reaction time sits between 100 and 120 milliseconds. This limit exists because of pure physiological hardware constraints. It takes approximately 13 to 70 milliseconds for a visual signal just to reach the brain. From there, your brain needs time to consciously acknowledge the input (around 75 to 150 ms) and send a signal down your spinal cord to actuate a muscle (another ~20 ms). Even with perfect anticipation, dropping below 100 milliseconds is physically impossible for a conscious, unpredicted response. I tried a number of things to optimize my reaction time. The ones that worked best for me were good sleep (8 hours uninterrupted in a well-ventilated room), weight training, testing when I’m neither too hungry nor too full, and doing it in a 22–23°C room with a high-refresh-rate monitor and a low-latency mouse. These helped me get close to my limits. I could probably push it a bit lower with caffeine, but I don’t use it because it makes me sick.
English
@ValorantBits I try to stay away from drugs. I’ve never tried them, so I can’t really comment, but caffeine did help.
English
@HCSolakoglu ADHD drugs improve it also and are not as harsh as caffeine
English
@mark_k @GoogleDeepMind I think 3.1 Flash will be released very soon. I’ve started running into A/B tests for it. Similarly, I’ve encountered a lot of A/B tests from OpenAI as well, so it looks like they might release a new model soon too.
English
A bit strange that we didn't get a Gemini release this week, after it was being hinted at.
Maybe it was planned but then @GoogleDeepMind had to postpone it in the last minute? 🤷♂️
Would have been nice to get Gemini 3.1 Flash, but maybe next week! 🤞
English
@FACEIT_Darwin I’m living for day Cache comes back and Mirage gets removed.
English
I’ve played 2400 FACEIT games… and only 17 were on Dust2 😅
What’s the one map you avoid because you think you suck at it?
FACEIT CS2@FACEITcs
Which map did players most want to avoid in February? Across 4.2M+ voting round 1 bans, Overpass takes the top spot (790,744), followed by Nuke (666,655) and Dust2 (610,226). Which map do you ban first?
English
Performance largely seems around the same after the recent update. Use my config files and you should be set!
English