Mike Sulka

Founder / Investor

“We’re trying to prove ourselves wrong as quickly as possible, because only in that way do we find progress.” — Richard Feynman

[Mike Sulka @SulkaMike]

AI policy isn't the fight. The people already understands the problem better than Washington thinks it does.

People don’t need AI Bosses writing or quoting another 70-page safety framework or a Senate hearing to convince them AI is capable. They use it. They see what it can already do.

Then they look at their own workday.

Email. Research. PowerPoints. Spreadsheets. Scheduling. Reports. Meetings about the meetings. Hours of administrative sludge that suddenly looks suspiciously automatable.

They understand the trajectory too. If AI can do part of my job today, what can it do in three years?

Then comes the simplest chain of reasoning in this entire policy debate:

AI runs in datacenters.
More datacenters mean more compute.
More compute means more capable AI.
More capable AI may threaten my job.

So when a giant datacenter shows up down the road, people don't experience it as an abstract discussion about American technological leadership.

They organize. Facebook groups. Town halls. Zoning boards. Elected officials.

Water. Power prices. Noise. Land use. Whatever argument has legal and political traction becomes the lever.

The industry calls it NIMBYism.

Maybe that's only the surface-level explanation.

The crowd has connected AI capability to the infrastructure making it possible, and some of them have concluded that slowing the infrastructure might slow what they fear is coming for their livelihood.

Not in my backyard.

What they're really saying is:

Not my job.

[Mike Sulka @SulkaMike]

I'm happy its Saturday! Friday saw a huge surge from team open source as Moonshot AI's Kimi K3.1 showed closed and costly's days are numbered. As are the days of a number of college football coaches. But hear this, security concerns continue to be super inflated.

Moonshot leads after a brief lull from across the Pacific. But they are shipping now! The new monster remains a monster: 2.8T total parameters, 104B active, 1M context, multimodal, and now $2/M input, $10/M output and 20 cents/M cached.

Now pay attention. Same size, 33% cheaper than K3. Moonshot promises full weights by November 4.

Artificial Analysis gave it 55, #4 out of 215 models, behind only Opus 5.5 at 58, Sonnet 5.5 at 56 and Muse Wonder Full at 56. AA also says 67% Terminal-Bench 4.0, $2.11 per Intelligence Index task and about 55K output tokens/task. Opus 5.5 costs $5.98/task and consumes 119K.

So not only intelligent, but significantly more token efficient as well. That's different. Many of the cheaper models have eaten their cost advantage with inefficient tokens.

The part almost everyone is going to skip. Moonshot intentionally made K3.1 worse at hacking before releasing the weights. Why would the closed community give credit when you can create a political controversy to protect your own product line?

Moonshot found cyber capability climbing with additional training. So for the open-weight version it removed cyber-related material from both mid-training data and reinforcement-learning environments while retaining ordinary coding. On Moonshot's internal exploit benchmark, the same pipeline without those reductions scored 90.5%. Final K3.1 Pro scores 81.0%. Old K3? 33.3%. The tasks include finding and exploiting real-style vulnerabilities in things like the Linux kernel, Node.js, JavaScriptCore and Exim.

Moonshot says it doesn't want open weights extending the absolute frontier of offensive cyber capability. Whether that strategy keeps working as general coding and long-horizon reasoning improve anyway is the question. And much a chicken egg discussion.

Meanwhile, OpenAI disclosed Friday that frontier model training no longer gets live internet access. What a concept eh ? Insert concerns and fears here. On Sept 20, OpenAI stopped its most capable training, hardened the environment and resumed on October 1 after an 11-day pause. Frontier RL now works against a huge indexed, cached imitation of the web instead. And for those paying attention. A concept Jensen Huang mentioned at GTC Berlin - an internet digital twin.

For the test, OpenAI pointed fleets of its best cyber agents at its own containment systems and told them to break out. Of course, they did because everyone knows that all security has holes. Persistence is a key red team capability, and if AI has one thing, its persistence and variation.

Security teams not using zero trust, and testing that trust are in for a world of hurt. The security game has been changing for a decade as speed from vulnerabilities to exploit has compressed. AI is just a continuation of that, this isn't new. What's new is that people less capable are now able to become active. Yet, here we are still standing, the world hasn't ended, a great depression and bank run isn't happening, the world goes on as IT security teams fix what they should have fixed anyway.

So if you are working overtime today fixing php, you have my respect. But clean it up by noon take in the Indiana v Michigan game, my MSU Spartans are waiting for basketball.

We've spent years asking how smart can we make these things? I get that some may be knee capping what they ship out of concerns about security, but the defenders need the best tools as well. Or there organization will get hit. But then again, organizations have been getting hit since NotPetya, Solarwinds, log4shell, wannacry... and the world didn't end.

Enjoy! Out.

[Mike Sulka @SulkaMike]

Happy Friday! We've detected a rumorish pulse! Three items are leaving fingerprints.

First, naming sanity. Kimi K3.1 is from Moonshot AI. It is NOT MiniMax M3.1 Flash, aka Space Bunny. Yes, China apparently decided we needed two hot 3.1 models in the same week. See if you can keep it all straight!

Moonshot is now teasing Kimi K3.1 itself, and Arena seems to agree. Our next anonymous model Kibitz has turned up in Arena tests, with tokenizer fingerprints and naming patterns pointing toward Kimi. Better than Random Bro analysis. Early Kibitz results are semi interesting as the 3D crowd persists. One Minecraft test had the suspected K3.1 place 13,707 blocks for 38 cents, versus GPT-6 Astra's 5,738 blocks for $5.31. If that held up at a real release... hmmmm.

You might recall, Kimi K3's biggest problem wasn't necessarily intelligence. A 2.8T-parameter monster that could be slow and difficult to serve. So K3.1 getting faster and cheaper may matter.

Qwen 4 says not so fast?

An Arena model called Kalmia is being fingerprinted as Alibaba/Qwen. Community tokenizer tests spin the wheel and it points toward Qwen. Alibaba has already talked about a Qwen4 family including Max, Flash, Plus and a 27B model. I'm anxious to see the 27B.

On the US shore, Microsoft merged a pull request into its EF Core repository where 17 evaluation configurations were changed from GPT-5.6 Terra to gpt-6.1-terra, plus one new configuration using it. Eighteen total. Terra is being used as the default judge model for those evaluations. Microsoft doesn't seem worried about dumpster divers. Terra historically occupies OpenAI's balanced middle tier. GPT-6.1 Sol already demolished the old value equation by giving us much more intelligence at dramatically lower cost than the prior flagship generation. If OpenAI has found another useful price/performance notch underneath it, fantastic. Bring it.

MiniMax M3.1 Flash weights are now live! But. The license is MiniMax Community, not Apache/MIT. Commercial users need attribution and notification, and products generating more than $20M/year need prior written authorization. Read your licenses, BEFORE you deploy.

MiniMax says Pro will come. Apparently a Bigger Bunny in the burrow.

What about Fable 5.5? LOL. Partners supposedly have it. Azure briefly exposed it. Tooling has left tracks. We've had “tomorrow,” “next week,” and “imminent” long enough that X influencers are in the grief stage of trying to get impressions.

This time next week we may look back and realize several of the next models were sitting right in front of us wearing fake mustaches. Til then, it looks like another quiet weekend unless I can talk Google into a Gemini Flash 4 release!

Enjoy! Out.

[Mike Sulka @SulkaMike]

Happy Thursday! Yesterday I wrote that AI isn't a GPU or even a server. It's an AI factory, and the weakest link eventually gets exposed. Jensen Huang and NVIDIA get that and still dominate, while I continue to miss that in the Warriors, despite Draymond, and Steph.

At GTC Berlin, NVIDIA unveiled DSX POD, essentially turning the AI datacenter into Lego set. Compute, networking, power, liquid cooling and cabling get integrated and tested at the factory, shipped in modular containers, then plugged into a prepared site. Fully absorb that for a moment. NVIDIA wants deployments to fall from roughly six months to as little as 14 days. Availabile early 2027.

Same number of tokens, 30% less power. Closed-loop cooling, water use near zero; at a 2GW deployment. 54 million liters savings of water every day (180,000 households). Vendor math, but... again, dominating the weak links.

Time to first token metric brings new meaning.

In the model world, Recursive Flint dropped. It's a 39.8B MoE coding model, 2.5B active NVFP4 - check. Apache 2.0, 262K documented context - good. Recursive reports 74.2% SWE-bench Verified, 89.7% SWE-bench Pro v2 and 54.9% Terminal-Bench 2.1. Benchies, good enough you might want to bring your own evals.

Built how? Recursive used a small human-and-agent team created Flint from scratch in under four months! Wow! They used its Eureka Machine to propose ideas, implement them, run experiments and decide what to try next. That sounds a lot like the overused, not understood term RSI or AI helping build the next AI. NVIDIA spoke and said, the specialized kernels improved Flint's RTX Spark generation speed 55% without changing the weights. And really grabbing my attention a freaking 220 tokens/sec on a 64GB RTX Spark. Recursive's went conservative at 150+ tok/s so any comments and proofs below would be appreciated.

Reddit already has it running on an M2 Ultra at roughly 40–55 tok/s with a Q5 quant, with one user reporting clean tool calls. Others say it gets stuck thinking or performs poorly on their code. So maybe let's not get too excited ??? Weights are on Hugging Face and it already has llama.cpp, Ollama, Unsloth, LM Studio, vLLM and SGLang support.

Quit scrolling!! OpenAI models are coming to Google Cloud. OK, you can scroll again! It matters, but mostly to those not built on AWS, or Microsoft Azure. SMB, Education. Google FTW first with all 3 major closed vendors. Enterprise AI will be more like picking the model that fits, flexible provider, rather than a devoted religious zealot.

Claude Duet! yawn. The egos at Anthropic have been so proud of Claude Code, they seem to have missed leading the charge making the tech more easily accessible. And yes, that is truly valuable to enterprises. Of course, being constantly computed constrained usually leaves Anthropic bragging without marketshare. Grok, Meta, OpenAI all beat them out of the gate.

I digress. Major model rumors, lots a chatter about not much. Even so, as NVIDIA's Vera Rubin full deployment rolls out earnestly (July/August), and could roll out even faster in 2027, that's the compute that will lead the next generation of AI breakthroughs... and dominate like my Warriors of 201x!

Enjoy! Out.

[Mike Sulka @SulkaMike]

I really enjoy the information's coverage. But this is much like describing the getaway car in detail, or the drill the robbers used on the vault.

Bad people will do bad things with good tools.

[Mike Sulka @SulkaMike]

Happy Wednesday! For those inclined to see AI as a giant bonfire of capex looking for revenue, SpaceX just slapped that argument in the face with a fairly large wad of cash.

SpaceX reported $14.58B of Q3 revenue, up 161% year over year, and $4.4B of net income. With the punchline being that Super Intelligence generated $7.56B, nearly 10X last year and 51% of company revenue. Yes, its true, AI is now a bigger revenue business inside SpaceX than Starlink. Cottage doomer industry be damned.

That being said, someone will soon annualize Tuesday afternoon into $3 trillion. Most of that AI revenue is selling compute capacity to customers including Anthropic, Google and Reflection AI. Not Grok subscriptions. SpaceX also spent $19.6B in capex during the quarter, much of it on compute infrastructure. It doesn't prove every AI datacenter being built is a gold mine. What it does prove is somebody is writing enormous checks for the compute. $7.56B in one quarter isn't an AI demo. It's a fabulous business most CEOS would die for.

SpaceX's AI compute capacity reached around 1.7 gigawatts, while Reuters says the company's main business currently looks closer to a cloud provider than an AI lab. And only people with a contrarian point of view would find smoke in that pivot.

Wemby and Shai aren't AI, so who wants to talk business during the NBA opening week! $SPCX did briefly trend overnight, so somebody noticed the money eventually.

DeepSeek always a crowd pleaser opened DeepInfer, the inference engine behind V4.1 Flash. The fun numbers are inside the production notes. One cluster serves 155 TRILLION tokens per day, with DeepSeek reporting $38 of revenue per GPU-hour at off-peak pricing. That's what cheap tokens look like after you multiply them by industrial scale. Further, DeepSeek says its 3FS system provides roughly 100 PiB of SSD storage just for KV cache. Agent cache hangs around three days; chatbot cache seven. Cut chatbot retention to three days and DeepSeek says cache-hit rate drops another 0.5%–1%. Cross-machine hits can even get pulled over a 400G VLAN.

There's a nice connection between these two stories. SpaceX shows billions of dollars flowing into AI compute. DeepSeek shows why operators are fighting over fractions of cache-hit rate once that compute reaches absurd scale. Models make headlines. Utilization gets money.

Not many rumors or models today. Mistral released Voxtral 3, a new speech and voice-agent stack. Mini transcription starts at $0.0035/minute, realtime at $0.008, and its bundled voice-agent API starts around $0.04/minute with Mistral Small 4 doing the reasoning. That's $2.40 for an hour of end-to-end AI conversation before telephony and outside tools. Public preview, but call-center economics continue heading in one direction.

More big techs report today. Meta, and Tesla! Toss in IBM for a thrill. But SpaceX owned Tuesday, Infrastructure wins again in the age of AI.

Enjoy! Out.

[Mike Sulka @SulkaMike]

Happy Tuesday! Cheap, fast and good. What happens when instead of only being able to pick two, we can pick all three?

Yesterday Space Bunny ate most of my attention, and I probably shortchanged Mercury 3. Inception is attacking the problem differently. Diffusion doesn't generate everything strictly one token at a time. These guys report 1,068 tokens/sec, 1M context, and a model more than 40X larger than Mercury 1 while largely retaining the speed. Cerebras type speed without the Cerebras architecture.

Plus a freebie, new API keys currently get 100 million free tokens, and launch pricing is 10 cents/M input, 35 cents/M output and one penny/M cached input. That's yet another version of price compression.

Buried inside Mercury 3's production web-agent harness worked on 50 live Mind2Web tasks. It completed 50% and, according to Inception, completed significantly more than GPT-6.1 Sol while using the same token volume and same number of agent steps. A production partner reports median model generation of 0.9 seconds and end-to-end responses around 2.2 seconds. 1,068 tokens/sec starts becoming more than a drag-racing number. It becomes instant.

Some early reaction is appropriately messy. One tester thought Mercury was surprisingly close to Luna on normal work; another watched it think for 12 seconds and dump an entire code file in another two or three, but wasn't impressed by the 3D result. Frankly, we usually aren't impressed by ANY 3D result. It appears worth a eval on your private set.

COBOL is suddenly a word you can say outloud in public again. OpenAI points out that COBOL still supports more than 40% of online banking systems, 80% of in-person credit-card transactions and 95% of ATM transactions. Seemingly, its not run with punchcards, who'd have thunk.

So why is it a talking point - one hint $. OpenAI's behavior-preserving COBOL migration test, GPT-5.6 Sol scored 5.4% at low reasoning. Yikes. Yet GPT-6.1 Sol hit 86.6% while costing 39% less, and high reasoning reached 97.9%. That's roughly a 16X improvement at the same low-reasoning setting. Private tests, but interesting.

Insert piles of money here. Thirty-year-old applications, retired developers, lousy documentation and business rules buried six levels deep in code everybody is terrified to touch because it still processes payroll every Friday. And somehow COBOL is trending on X this morning, sitting between Vogt and Cosmi on a board dominated by football, baseball and WWE. Grace Hopper's asking who's buying.

Remember CPUs? Modal got NVIDIA's new Vera CPU and built a benchmark around the ugly work AI-agent sandboxes actually do: compilation, grep, git operations, device emulation and containers. They used 50 Terminal-Bench oracle solutions, stripped out the network calls and ran them concurrently. On a quiet machine Vera was good, not magic. Then Modal piled on the jobs. Under contention, Vera delivered up to 2X the task throughput of an equivalent Intel Granite Rapids machine, while individual latency stayed equal or sometimes lower. The competing CPUs eventually flattened out as jobs accumulated.

Vera kept scaling. Modal says the key is memory bandwidth and the ability to pack more container workloads per core without throttling. The hardware is 88 Olympus cores, 176 threads, up to 1.5TB of LPDDR5X and roughly 1.2TB/sec of memory bandwidth. But here's the thing, Modal points out that during reinforcement-learning rollouts, expensive GPUs can literally sit idle waiting for sandbox jobs to finish before the next training step starts. Make compilation, testing and filesystem work faster and suddenly your CPU improves GPU utilization. Most miss this.

Remember AI isn't a CPU or a GPU. AI also isn't just a server either. Its a rack of servers and systems that integrate with a row of racks and those rows all work together to form the proverbial AI factory. The weakest link gets exposed.

Agents have been dragging the rest of the datacenter back onto the field. Sort of like Kittle blocking downfield while everybody watches the quarterback, the flashy guy gets the highlight, but somebody still has to make the damn play work. Vera apparently is ready to make the damn play work.

Useful AI needs great infrastructure even if a diffusion like breakthrough emerges. I've talked alot about cost in this space, because speed breakthroughs haven't been exactly breathtaking. But the industry does keep taking steps to where we are getting on intelligence and cost.

So I ask again.. Cheap, fast and good. What happens when instead of only being able to pick two, we can pick all three?

Enjoy! Out.

[Mike Sulka @SulkaMike]

People continue to conflate"frontier performance" (scoring high on a test) with "general intelligence" (the ability to function reliably in the real world).

[Mike Sulka @SulkaMike]

For a company working towards mass adoption, completely agree optics is the real issue since Google somehow thinks finance is best option. Just not sure it's a current or future issue. But it will grow as Waymo does.

But it's a sad state of affairs that 100% tarrifs don't at least level the playing field for a US builder. Anyone??

[Mike Sulka @SulkaMike]

Motivations are tough to address when the mem line item for the GPU is staggering. Maybe customers are pushing Nvidia. But if the big 3 mems go along with this NVHBM base-die standardization on TSMC 3nm… not sure if they really have a choice or if they can walk away. And if they don’t sell it to NVIDIA, who else do they sell it to at that volume? All the leverage seems to move to Santa Clara HQ.

[Mike Sulka @SulkaMike]

Happy Monday! MiniMax M3.1 Flash has arrived! Giving Space Bunny a new name, and revealing this bunny was busting out the tokens, more than 200 TRILLION tokens! That's one hell of a focus group.

Find it on MiniMax API, M Plan and MiniMax Code. One million token context, text/image/video input, five reasoning levels from low through max, and MiniMax says output can exceed 200 tokens/sec with median speeds above 170. Weights not out yet, but coming.

For those bench based, MiniMax reports 73.7% DeepSWE, 61.7 OSWorld V2, 29.6 Agents' Last Exam and 82% CyberGym. Nice numbers. Third parties should weigh in shortly. Costs, 15 cents/M input. 60 cents/M output. 1.5 cents/M cached input.

Now compare that with our new DeepSeek 4.1 Flash 1017 friend. DeepSeek same 15 cents/$0.60 for uncached work during off-peak hours, but 2x that to 30 cents/$1.20 during peak hours. That's the cash, but the cache still leans DeepSeek's direction. At 0.3 cents/M cached tokens off-peak and 0.6 cents/M peak. Either way, close enough to let the work drive the model selection.

Reddit called it. Tokenizer tests tied it strongly to the MiniMax family, and one user reported Space Bunny at roughly 85–100 tokens/sec versus 82–112 for M3.1 Flash Preview.

But X seems to be meandering past this one talking about the Cowboys, Good Monday, Goku, PC gaming and Jerry Jones. AI hasn't cracked the top ten for the day. Show the Bunny some luv!

But well armed, check this out! SGLang 0.5.22 landed Monday morning with 950 pull requests from 245 contributors. Deep in the tech notes, SGLang says DeepSeek V4.1 now gets up to 2.3X more throughput per H200 GPU. Long-prompt prefill/decode disaggregation gets up to 27% higher throughput while cutting p95 time-to-first-token roughly in half. GLM-5.3 gets up to 2.1X faster decoding.

This is where real works gets done. Infrastructure geeks.

Tack on , SGLang integrated InstantTensor, which it summarizes simply as: Models load 10X faster. Wow!
Yes, Qwen3-30B-A3B on one H200 went from 57.4 seconds to 1.77 seconds. My calc says that's an undersell and 32.4X faster! That being said, vendor reported, but I guess they have some room during the test. Like an F1 driver pleasing a road driving license test.

How about another infrastructure lesson? OpenCode says a security breach related to what it calls a large-scale distillation attack affected 1.4% of Console users with external model providers connected. The intruder apparently reached a legacy system OpenCode was already decommissioning, and third-party provider keys stored in those workspaces may have been exposed. An an old system holding API credentials.

Thirty years in IT and eventually somebody rediscovers every lesson is if it is new. And be careful WHO you store your keys with, and what default billing maxes you use, and smart people don't auto recharge their API keys or hold on to them for extended periods of time. They budget, fill and rotate. Yeah, that's not the easy button, but it works from a security stance.

But my side gig is sure interested in this, IBM Research has AI agents designing classical hardware used for quantum error correction. Nice! They coordinate via an agent. It assigns ideas to four worker agents. They change the hardware implementation. A verifier checks that the changes still work. Hardware synthesis reports the latency and resource use, and the system loops again. The all-open-model version explored 100 different hypotheses and produced a design 52X faster. Nice work IBM!

So Happy Monday is off to a good start, Elon says more SpaceX launches coming, maybe even tomorrow! Now let's see what you can do to change the world starting today!

Enjoy! Out.

[Mike Sulka @SulkaMike]

You want some more on DeepSeek, seems many do! Yesterday I wrote about DeepSeek-V4.1-Flash-1017, and a bunch of people with too much hardware and not enough sleep started unpacking.

One dev compared the new weights with the September checkpoint and reports that approx 40% of routed expert tensors changed, 30% of the Engram tensors changed, and 50% of the vision tower’s expert tensors changed. Same architecture. So one view is that this was “major revision.”

Another dev has 1017 running across two DGX Sparks using a 2.9-bit EXL3 quant. Same machines, same weights and same prompts, TensorFold delivered 56.8 tokens/sec on text versus 27.0 with the vLLM kit, and 92.2 tokens/sec on code versus 39.9. Wow. They listed other context ranges, with the same story.

Twice the performance.

Another DeepSeek number from Reddit says that a Nebius engineer ran 1017 through a personal coding benchmark built from 10 real tasks, 3 runs apiece with hidden tests. DeepSeek at high effort passed all 30 runs for about 4.5 cents per task. Excellent. But GPT-6 Luna did the same for 1.45 cents, Haiku 5.5 for 2.08 cents, while Opus 5.5 low cost 11.68 cents. But the crazy number is that DeepSeek consumed around 550,000 tokens per task. Opus used around 30,000.

There are your cache wars again. DeepSeek is so cheap per token that it can eat nearly twenty times as many tokens and still come out cheaper than Opus.

And another test shows something similar. DeepSeek produced a respectable result for about 23 cents versus $6 for Opus, but DeepSeek took around an hour versus 20 minutes. Yet another, DeepSeek took about 90 minutes and cost 14 cents, while Haiku cost 8 cents and finished much faster.

We are redefining good and cheap and the analytics may or may not catch these nuansances. Are cheap tokens, cheap work. Maybe.

And people were talking... In one U.S. X snapshot at 14:00 UTC, DeepSeek was #31, while Buckeyes, Caturday, Johnny Carson, Steve Kerr and various college football terms bounced around the list. DeepSeek showing up one slot ahead of Steve Kerr on a college football Saturday feels about right.

For today, there was no second Saturday whale hiding in our seas. Fable 5.5 remains "soon". Qwen 4 remains “soon” with the best current end-of-October claim still basically In Bro We Trust. Kimi K3.1 remains smoke.

Meanwhile, IBM quietly shipped something fascinating over a weekend. I guess I thought these guys were M-F.

Enter Granite Switch 4.1 which weighs in at 8B and 30B. A general-purpose decision adapter, a specialized Guardian adapter and a 470M speech-recognition model into one system. IBM says the whole combination occupies approximately the same GPU VRAM as the base model alone. Activated LoRA adapters reuse the base model's existing KV cache instead of building another one.

Chat model. Decision model. Safety classifier. Speech recognition. One endpoint. Essentially one base-model memory footprint.

IBMers even demoed a custom 3B version making 60–100 decisions per second while playing Doom, versus a reported 10 decisions/sec for Jev. The Granite model could still talk about what it was doing. More importantly for people who aren't playing Doom on the company GPU cluster, Switch runs through vLLM and now has llama.cpp support.

Mistral also released the nonsexy Timestral. It isn't another chatbot. It's a family of time-series foundation models ranging from 5 million to 209 million parameters. Think equipment vibration, temperature, pressure, capacity, energy demand, inventory, financial flows and observability metrics. The largest 209M model currently leads Mistral's reported results across GIFT-Eval, TIME and fev-bench, while the 123M version comes within 0.2% of the largest model's score.

And heaven forbid, it doesn't need a GPU. AI watching industrial telemetry and potentially telling somebody a pump is wandering out of tolerance before it craps itself. No glowing voxel castle required.

Ai2 managed another useful-AI number. A new HiRO-ACEv2 climate system generates global climate data at 3-kilometer resolution for roughly 1/300th the cost of the X-SHiELD high-resolution atmospheric simulation it emulates. Extreme rain, wind and pressure events are no problemo. Optimize and drop the cost. 300x. Let's see OpenAI, or Anthropic do that!

Also, Hugging Face quietly put llama.cpp in the Windows 11 system tray. Llama Windows can download models from Hugging Face, launch and manage llama.cpp without a terminal or administrator rights, expose an OpenAI-compatible API on localhost, share the normal Hugging Face model cache, and already supports Windows on Arm and NVIDIA RTX Spark. That may put local AI in front of a hell of a lot more normal people, but then again, do you look at your Windows taskbar for news and weather??

Its been a crazy week in Open and optimization. Smart humans keep figuring out how to make them actually useful, and cheaper to run. What will this week bring?

Nothing is 100 percent, but I'm excited!

Enjoy! Out.

[Mike Sulka @SulkaMike]

Well damn. Yesterday I wrote that China had gone quiet for 24 days. My friends at DeepSeek took that personally and boom! Just hours ago, DeepSeek drops V4.1-Flash-1017. A quote “minor update” to Flash.

Minor, my foot. Which explains why my inbox was filled with DeepSeek chatter. With Daniel and "DeepSeek is the goats" first comment in! LOL.

Details, let's talk about this a bit, DeepSeek kept the same deepseek-flash API endpoint and the same pricing. And there is a pricing story no one is talking about, but I'll detail at length soon. Benchies say Terminal-Bench 4.0 jumping from 31.2 to 42.3, ProgramBench from 20.3 to 29.1, HLE from 36.8 to 42.6 and DeepSWE from 74.2 to 77.0. Plus the new checkpoint finishes work faster and with shorter outputs, not sure Visa measures sub penny daily charges.

Reddit's first reaction, like mine. DeepSeek calls it a minor update, then publishes a table where it suddenly beats Sol and Opus 5 on some of its chosen coding and automation tests. But, be careful, these are DeepSeek's benchmarks, third parties will weigh in shortly. But DeepSeek hasn't been a hype machine, they have been a drop machine.

DeepSeek also opened its internal Search and Fetch APIs to everyone. Their search engine reportedly handles hundreds of millions of searches a day, indexes hundreds of billions of pages, supports multilingual retrieval and specifically claims to filter SEO spam, GEO manipulation and low-quality AI-generated content. Search returns structured results; Fetch turns a webpage into Markdown for an agent.

The price? 75 cents per 1,000 successful searches. Another 75 cents per 1,000 fetches. So 1,000 search-and-fetch pairs cost $1.50, that's before token burn using those results. Yet, for the moment, server-side Search/Fetch calls used with DeepSeek and its Harness, Claude Code or Codex are free until the end of October. Something many will miss.

DeepSeek is starting to sell the plumbing agents need around the model. Search, fetch, Responses API compatibility, tool calls, million-token context and an Anthropic-compatible endpoint. You might call that an agent infrastructure provider.

And the entire weekend off-peak. Today Flash costs $0.15/M for uncached input, $0.60/M output and a ridiculous $0.003/M for cache hits. Peak prices during the workweek are double.

So yes, even Chinese inference celebrates college football Saturday. While #4 Ohio State and #5 Indiana are getting ready for noon, DeepSeek is running a half-price happy hour all day to allow me to press enter during commercials. Toss in Alabama-Tennessee this afternoon and Notre Dame-BYU tonight, and I might be able to rebuild the entire Adobe suite of products. Oh wait... someone already did that. And If Qwen 4 lands between the second and third quarter, I'm blaming yesterday's post.

Just a quick Friday update, Reflection delivered Beam's actual weights. We talked about the 501B-total / 23B-active model when it was announced, but Friday matters because the promise became a download. BF16, FP8 and NVFP4 weights are now up, along with the model card, safety report, speculative-decoding drafter and Mirror coding harness. And unlike Meta's Muse Spark Open license adventure Thursday, Beam is Apache 2.0.

Now Beam is still 501B. So “have fun” assumes you own enough unified memory to heat a small church. But 23B active is interesting, NVFP4 support is real, llama.cpp support has already merged, and Reflection says Beam uses roughly three times fewer reasoning tokens than comparable open models.

Its been an amazing week. Meta shipped an 816B American open-weight model. NVIDIA shipped local-friendly Nemotron models. Reflection released a 501B model under Apache 2.0. Then DeepSeek ended the Chinese silence with an MIT-licensed update and opened the search engine underneath its agents.

And while everybody stares at Fable 5.5 fingerprints, they gave Claude Mythos 5 eight quantum-circuit optimization problems and an external evaluator. Claude would design a circuit, run it, get a correctness check and score, then revise the design without somebody steering every iteration. Across all eight problems, Anthropic says Claude found designs using fewer qubits and fewer gates than previously published approaches. One resulting design uses 12% more qubits but 17x fewer gates than the smallest previous design. Now we're talking.

The model proposes. Something independent checks the work. The system iterates. Claude didn't wake up with a quantum physics PhD halfway through the run. Anthropic surrounded capable intelligence with a measurable objective, an external verifier and enough time to work the problem. Useful AI again.

Rumors do continue, but Fable 5.5 still hasn't shipped. GLM 5.5 and Kimi K3.1 are again rumors, not shipped.

What did ship, DeepSeek! So build your work plan, press enter, set your loop to autonomous mode and let it cook while you enjoy the kickoff, and then another kickoff, and another kickoff!

Enjoy! Out.

[Mike Sulka @SulkaMike]

@sama @elonmusk I've enjoyed mine! How much $$$$

[Mike Sulka @SulkaMike]

@harden_369741 I have plenty of bruises from the keyboard showing the school of hard knocks seems to be my path more than my vram calculator.

[Mike Sulka @SulkaMike]

Pumped for Friday after a American open-weight Thursday. Meta shipped. NVIDIA shipped. And Microsoft teased! Curiously 24 days.

Last major Chinese release I recall is Xiaomi MiMo-V2.6, released September 22, 2026. 24 days ago.
DeepSeek V4.1 Flash was September 10. Qwen3.8-Max was August 3. Beyond that, a few image models, harness updates, and smoke but no new Chinese heavyweight chat release. China has usually quiet. Hmmmm.

But America Open is rocking! Meta's Muse Spark Open is real and the weights actually shipped. It is a big boy: 816B total parameters, about 42B active, multimodal, million-token context, so your DGX spark is going to need some friends. Artificial Analysis scores it at 39, it is now King of the U.S. open-weight models. Its sure nice to see Meta back in this space.

To compare, GLM-5.3 is sitting at 45, Kimi K3 at 44 and Qwen3.8 around 40. Remember though intelligence is just one measure these days.

Apple and Meta. Not exactly besties, they usually are kicking each other in the shins over privacy and advertising. Now, Apple optimized its MLX stack to distribute Muse Spark Open across M5 Ultra Mac Studios using RDMA over Thunderbolt. So it fits on a desktop, right? Well... not exactly. Apple says the NVFP4 model itself occupies about 493GB, and once inference working memory is included it won't fit into a single 512GB Mac Studio. But for those with two 512GB Studios, have we got a deal for you!

Four M5 Ultras push decode speed to about 36.7 tokens/sec. The Wall Street Journal puts the required 1TB-class Mac setup north of $40,000.

Technically, it works on local. Technically.

Reddit's reaction was basically, cool weights, where am I finding 700GB of VRAM? Then somebody discovered the two-Mac requirement and the conversation became considerably less Walmart and considerably more data center.

Another gem hiding in the fine print. Muse Spark Open is open-weight, but this is not Apache 2.0 throw-it-over-the-wall open. Meta wants so love. So you'll be displaying "Built with Muse Spark Open." And if you're offering models as a service and your business exceeds $20 million of annual revenue, you need a separate agreement with Meta. Open weights - check.
Lawyer - check.

That should have been enough for Thursday, but Nemotron 3.5 Bolt dropped too. 74B total / 9B active and 24GB using native 2-bit quantization. Plus llama.cpp, Ollama, LM Studio, Unsloth and vLLM, day zero. NVIDIA explicitly designed this one for RTX systems and DGX Spark. Artificial Analysis only gives it a 21.

Perfect direction. Dense 70B crawls to a stop on local eq. Its scheduled for testing against by normal local pal, Qwen3.6 35B-A3B (reasoning) which checks in at 18 on AA. But not there for Qwen3.8 in raw intelligence. But, the direction is good. And like me, the community sees the combination of 74B total, 9B active and 24GB actual footprint as good for local.

And golly pa, look who else is serving on Thursday, well, not quite, but... Microsoft walked but the serving wasn't exactly something you take home.

MAI-Base-2 is 2.8 trillion parameters / 80B active, trained across 16,000 GB200s for 38 days on a 50-trillion-token mixture. Microsoft says its training recipe is 4–12x more FLOPs-efficient than DeepSeek-V4-Pro-Base and that MAI-Base-2 beat every open-weight model it tested on its private held-out pretraining evaluations. Some bold claims Redmond.

But, its a tease. Base model is currently entering post-training. The Microsoft line most missed is probably not 2.8T. Microsoft says owning its model pipeline end-to-end lets it optimize models and infrastructure together, and it plans eventually to license weights to trusted Foundry partners. That's Microsoft talking like a company with options when the contract negotiations begin.

But you see, there's a Chinese connection just when you didn't expect it. Microsoft explicitly used DeepSeek, Kimi, Qwen and MiMo as comparison points. Meta benchmarks Spark Open against the Chinese opens. NVIDIA's Reddit comparisons immediately turn into Qwen discussions. Yet China's last major new model was September 22. 24 days ago.
.
Maybe a quiet release cycle. Maybe DeepSeek V4.1 Pro, Kimi K3.1 and Qwen 4 aren't ready. But after hurricane of landfalls, a month of silence while U.S. companies suddenly unload open models seems odd.

Ah, patience again.

Fable 5.5 is of course rumoring. Anthropic EAP names resembling Onigiri and Gabagool, Azure/DNS footprints and reported silent routing are all chatter bound.

If Friday could top that, we should have started at midnight, otherwise, we have some new stuff for the weekend once we fake work through Friday!

Enjoy! Out.

[Mike Sulka @SulkaMike]

@harden_369741 LOL! Even when you have experience seeing the model parameters, and actives, you still just have to try sometimes!!

[Mike Sulka @SulkaMike]

Under the hype PR machine, Anthropic had a good quarter. A really good quarter. The real story is impressive enough without turning $16.8 billion into $100 billion with a calculator and a dream.

Anthropic told prospective IPO investors that Q3 revenue reached $16.8 billion, up 45% from $11.6 billion in Q2 and $4.7 billion in Q1. Quick calc and Anthropic has generated about $33.1 billion through the first nine months of 2026. A number just about any company would die for. If Q4 repeats Q3, the company finishes the year about $50 billion. Another stronger quarter might put 2026 revenue somewhere closer to the $50–60 billion.

Meanwhile, social media is talking about $100–110 billion. Yeah, not exactly. But of course, its predictable social media. Shit posts, and BS.

You could say the disconnect is run rate. Analysts are talking about Anthropic potentially exiting December at a $100–110 billion annual revenue run rate, but that's not a real number. Kind of like Best Buy using Black Friday sales to indicate their annual take home.

Plus profits. Anthropic told investors it was profitable on an adjusted basis, reportedly for the second consecutive quarter. Bloomberg's reporting also has it generating positive free cash flow, with an adjusted gross margin around 85% and adjusted operating margin above 11%. Now that being said, those are private-company numbers presented to prospective investors, not a nice boring public 10-Q legal document.

Even so, credit to Anthropic. $50B and that's not parameters.

Reddit played with ARR, half the crowd multiplied $16.8 billion by four and announced $67 billion ARR, somebody else pointed out that ARR normally uses the latest monthly run rate, and eventually everyone remembered that none of those numbers are the same thing as full-year revenue. Finance remains undefeated at ruining a perfectly good engagement headline.

But for the techs, Anthropic also shipped something Wednesday: OpenTelemetry tracing for Claude agents. Model calls, tool use, MCP calls, latency, token consumption and IDs can now flow into the observability tools enterprises already use. Anthropic's own recommendation is basically: trace the agent, watch what it does, then expand its access as it proves trustworthy. Bout damn time.

So its good to see an AI company besides Google actually using their tokens to build something around auditing and identity. That is how this stuff gets deployed. Not because the CEO watched a demo where Claude built a glowing 3D castle, but because operations can answer: What did it do? Which tool did it call? How long did it take? How much did it cost?

While OpenAI's 28-day bonanza fizzled with a reset, Anthropic piled on with idle-compaction work in Claude Code is another little piece of the same puzzle. Long sessions can compact before the prompt cache expires rather than letting a huge context go cold and paying to ingest the whole mess again later. Plus a switch to disable it.

Which leads right back to Cache Wars. The sticker price of the model matters, but increasingly the harness determines the bill. Cache expiration, context growth, compaction, subagents, routing and tool calls can turn the same model into either a bargain or a budget meeting.

Space xAI, or is is SpaceSI, or XSI, who the heck even knows.... said Grok Voice 3.0 is answering about 20,000 Starlink customer calls per day, with roughly 70% resolved without a human. One in four sales calls reportedly ends in a purchase. The voice agent can verify accounts, troubleshoot hardware, issue credits and ship replacement equipment. It costs $0.08 per minute.

The big-model rumorboard itself stayed mostly where it was. Fable 5.5 has more Azure/DNS smoke, including reported infrastructure dating back to September 28. Nvidia is teasing something new for this week, while CatJam and Wild MissingNo remain anonymous fast models rather than anything I'm calling a frontier release.

The Chinese labs remain unusually quiet. Hmmmm. If only my DMs could talk.

Enjoy! Out.

[Mike Sulka @SulkaMike]

Hallelujah, useful AI!! Argon escaped Fairwind. Google talked about 300 TiB, 800,000 lines of code, and 2.7x.

While we were waiting, Google used Argon agents to analyze telemetry across its datacenter fleet and found memory optimizations that have already freed more than 300 TiB, with another 500 TiB to 1 PiB potentially available. That's real money. Smart Infrastructure Managers would replicate this in their environments. That’s the kind of AI result enterprises understand. They do NOT care that Argon beat Model X by 1.7 points.” They DO care about finding a warehouse full of memory you were already paying for.

Google used Argon to migrate C/C++ code to Rust, including work touching more than 800,000 lines of Fuchsia’s Zircon kernel. Another project replaced 32,000 lines of SIMD code in the Rust port of libgav1 and ended up 2.7x faster while producing identical output. And for those concerned about the humans, yes humans still automated testing, still reviewed the work before it gets near production. Change control mode is still on.

While nearly every AI model company can build a frontier model these days, Google is showing what happens when useful is valued.

Of course Reddit immediately supplied balance. One user gave Argon a PDF with upcoming flights and asked it to add them in Google Calendar. It failed. Claude finished the job. I get it. I see these things happen with Gemini Spark as well. But, if you build auditing into the process, with a redo button, you'll have a happier life.

People have been saying some version of this since the first developer told us, “It works on my machine.”

Speaking of infrastructure, Liquid AI dropped LFM2.5-24B-A2B, a 24B MoE with only about 2.45B parameters active per token, 131K context, and support for llama.cpp, MLX, vLLM, SGLang and ONNX out of the gate. Wow, say that 3 times fast!! Liquid reports 184 tokens/sec on an M5 Max and 114 on the Ryzen AI Max+ 395.

That architecture is exactly what I keep asking the local model world for. Enough total capacity to be useful, very few active parameters so it doesn’t move like a Buick towing a camper, and small enough that a DGX spark can blast out tokens per second.

Liquid says it competes with larger MoEs on several agent and tool-use tests, but Qwen3.6 beats it on knowledge and coding. Imagine that: a benchmark chart containing losses. I like the direction, but remember to eval with your standard test set before clicking default.

Good grief drama queen OpenAI was in the news with a story that most companies would have kept quiet. Federal prosecutors allege an OpenAI engineer placed an encrypted audio recorder behind a conference-room television, plugged it into USB and hid it in toilet paper. Good ole human problems. Spy v Spy.

The AI industry spends years talking about AI worms, autonomous cyber agents, model replication and attacks running at machine speed. Meanwhile one of the most valuable AI companies on Earth allegedly had an insider-security problem involving a television, a USB port and bathroom supplies. There’s your Thursday security lesson reminder. Remember when Raspberry Pis being plugged into open switch ports on unsecured network closets was a thing?

Physical access, least privilege, cameras, removable media, conference rooms and occasionally checking behind the damn TV did not stop mattering because somebody invented superintelligence.

So when model releases heat up, the alternate fan boys start yapping, me too!! Fable 5.5 was said to have showed up in Azure-related listings. GPT-6.1 Astra chatter remained chatter. Nano Banana 3 Pro made the rounds. Carbon is still an internal Google checkpoint. Nothing material moved on DeepSeek V4.1 Pro, Kimi K3.1 or Qwen 4.

So thank God for Google! Handing out practical mature lessons on how to deliver value with AI!! AI keeps getting more exotic. The smart humans figure out how to make it work in real world cases.

Enjoy! Out.

[Mike Sulka @SulkaMike]

The Watermelon finally ripened. The model we've been tracking for weeks actually shipped Monday as Muse Wonder and Muse Wonder Full. So scratch Watermelon off the rumorboard. This one is now green, and ripe.

Meta says Wonder is its first real frontier-class model family, built around personal and professional agents, with Active Memory and Active Focus for remembering your life and grinding through long-running jobs. Wonder is live in Meta's API and Muse Code; Wonder Full, the larger version, is restricted to Muse Code subscriptions.

It also seems to be good. Artificial Analysis has Wonder Full at 56, only two points behind Opus 5.5 at 58 and exactly tied with Sonnet 5.5. Ordinary Wonder comes in at 52, tied with GPT-6.1 Sol. It seems much better than having to explain Llama 4 results.

And... Meta expects you to pay for this one. Charging $4/M input, $20/M output and $0.20/M cached input for regular Wonder. Pricing that implies they have a winner. Yet, Artificial Analysis measured what the model actually eats. Wonder used about 7.69 million input tokens per average Intelligence Index task, and AA calculates $4.52 per task. GPT-6.1 Sol scores the same 52 and costs roughly $0.72 per task.

Meta gets spanked, or AA has some explaining to do.

The hidden numbers explain that roughly 76% of Wonder's AA task cost came from input, not output. Even $1.92 was cache writes, $1.44 cache hits and only seven cents uncached input. Meta gave Wonder a 95% cache discount and still managed to build a model whose context appetite can eat the savings.

Reddit found the same problem using less scientific terminology. Early Muse Code users are complaining that Wonder Full can devour subscription limits during long jobs. One reported a task that consumed about 21% of a $50 plan's weekly allowance; another simply described the usage as the problem.

Due to historic activities, the more amusing Reddit debate is whether Wonder is benchmaxxed. People threw lava lamps, Minecraft clones and front ends at it and immediately declared civilization had been deceived when some outputs looked mediocre. Despite these non Enterprise use cases, Meta's own people stepped in and pushed back that they didn't heavily RL the model on 3D game generation.

But maybe the gamer crowd does lead to a legitimate question: how well does your benchmark winner generalize to something its post-training team didn't specifically care about?

And Watermelon is proprietary. Wonder is sold through an API at Opus-class prices. The best version, Wonder Full, isn't available through the API at all; you have to consume it inside Muse Code. Meta is selling what it calls "personal superintelligence for everyone," but the flagship version currently lives behind a $50-200 coding boundary, that even a progressive teen age allowance isn't going to hit.

And... Muse Code subscriptions default to Contributor mode. In exchange for more usage, Meta can use that content to train future models. You can turn Contributor mode off. But you know the drill.

Another tiny contradiction I love. Meta's pitch centers on Active Memory and Active Focus: the agent remembers you, understands your finances, health information, legal documents and personal preferences, and keeps working over time. Meta says Muse conversations aren't shared with its advertising systems, lets users control connected services and provides audit trails. That's good architecture. But the smartest agents still need somebody else's website. The WSJ reports Amazon has already blocked Muse from completing some transactions. Personal superintelligence, meet the TOS.

Meanwhile, Reuters reports OpenAI and Anthropic are negotiating with supposedly independent model evaluators over what those evaluators can publish. Both companies publicly support independent evaluation, but discussions include restrictions around sensitive results, trade secrets and whether evaluators can work for competing labs.

So much for the "independent safety evaluation" stake, the aforementioned pounded into the ground as good policy.

An independent auditor whose final report requires negotiations with the company might be useful. But accountants have another word for management approving the audit narrative.

Nothing material moved overnight on Carbon, Bel, Fable 5.5, Grok 4.8, DeepSeek V4.1 Pro, Kimi K3.1 or Qwen 4. OpenAI's 28-day campaign hit Day 8 and delivered another global usage reset rather than another anything.

Which leads me to.... a reminder that while Meta delivered something. What Meta didn't do is deliver on the promise of open.

Enjoy! Out.