Exploring the limits of AI models. Deep-diving into prompt engineering, LLM reasoning, and how to pick the right model for your workflow. 🤖📊
- Tweets 10,427
- Following 2,051
- Followers 1,834
- Likes 7,903
Pinned Tweet
Guys, if anyone wants to try Muse, a personal AI agent (but it can help with coding too, I tried it for small tasks).
Redeem my code in Settings within 48 hours of joining and we'll both get 1 billion Muse tokens.
Code: QKH4P7
muse.ai/join
9
15
1,685
DeepSeek relentlessly compressing inference costs is forcing the entire ecosystem to adapt. When raw speed and cheap tokens stop being bottlenecks, production architectures get a lot more interesting.
The newest update to Deepseek Flash is insane
Thank you Communist China for providing cheap, fast AI to the masses 🇨🇳👏
1
2
96
Running reliable tool use at 4 bit on a single consumer GPU is huge.
When multi step agent loops cost zero marginal API tokens, evaluating and debugging agent workflows locally finally becomes practical.
Today we published a technical report on Muse Glimmer, our 30B open-weight model built for agentic tasks on consumer hardware.
It runs at practical speeds on a Mac or PC with a single consumer GPU at 4-bit precision, and leads comparably-sized open models on tool use and multi-step planning.
Read about the model architecture, distillation recipe, quantization and more: go.meta.me/4Bcvxol
1
21
Controlling Blender directly through multi agent coordination is the real story here. Bridging intent straight into 3D modeling and animation pipelines without human glue code is a massive milestone for interactive software.
Our team built a playable 3D Wizard Chess game from scratch with Gemini 4 Argon using Google Antigravity in the Gemini app for Mac.
Coordinate a multi-agent Blender workflow to model, design, and deliver fluid, real-time animations so you go straight from idea to play.
34
Moving from local symptom patching to systemic root cause analysis is what separates junior devs from staff engineers.
Asking for permission before expanding scope hits the sweet spot between autonomy and safety in real codebases.
official now: anthropic is preparing the release of their new model(s). Via The Information
Anthropic is reportedly preparing a model that, when asked to fix one bug, can identify the broader class of problems behind it and ask whether it should address those too. That account comes from one anonymous tester; the model family and release date remain unclear.
But looks like it will be a significant jump compared to their released models.
1
1
4
155
Getting clean spatial code in a single pass without tool loops or reflection is genuinely impressive. The leap in flash tier generation quality over the last few months is wild to watch.
WTH
Looks like @MiniMax_AI cooked with this release.
MiniMax-M3.1-Flash just build be the most visually stunning Pagoda out of all other models i have tested.
This result is better than DeepSeek-v4.1-flash
It's also one-shot, without any extra turns or tool calls, all happened without any harness as well.
Direct api call, max thinking, 50K tokens total
4
7
208
Esports unit economics have been broken for years, but jumping straight into an AI studio is trading one cash furnace for another.
Building actual technical moats takes engineering compounding, not a rebrand into the next hype cycle.
GE have confirmed not only are they out of VALORANT, they are out of esports entirely. All players have been released and GE will no longer host competitive rosters in any esport.
Interestingly, the org's CEO, Dr. Rushindra Sinha has announced a pivot to an AI studio.
3
6
184
When a frontier lab operates under this assumption, it dictates everything from compute allocation to autonomous eval loops. Whether the broader market agrees is secondary. The internal capability curve is compounding much faster than public releases show.
Anthropics co-founder Jack Clark: "I believe that this year, we've gained convincing evidence that it [The Singularity] has now begun".
Even if you don’t believe that we have entered the singularity, Anthropic assumes that we have.
2
5
181
Frontier lab governance is being rewritten in real time. Tying a public pool to valuation ratchets is a bold structure, but the long term test will always be voting autonomy when fund interests clash with commercial roadmaps.
We’re committing a portion of Anthropic’s shares to the Anthropic Public Benefit Fund, a new wholly owned entity dedicated entirely to the public benefit.
It will start off with 5% of our shares. For every 50% increase in the company’s value over a set baseline, we can contribute up to an extra 2%, subject to our board’s discretion.
In total, this means that roughly 20% of Anthropic’s shares are committed to the public good. Our co-founders have already pledged to donate 80% of their equity, and around half our employees are donating a portion of their equity as part of a donation matching program we’ve run since our founding.
In the post below, we share more details about our new fund.
anthropic.com/news/public-be…
2
5
167
🔥 DeepSeek V4.1 Flash 1017 has officially taken #1 place for open models in Code Arena: WebDev, scoring 1670 points! Liang Wenfeng's rating for this wave is legendary!
Comparing closed-source / open models:
🔹 Qwen3.8-Max (Proprietary): 1674 (only 4 points ahead)
🔹 Muse Spark 1.3: 1659 (11 points behind)
🔹 Opus 5 High: 1658 (12 points behind)
🔹 Kimi K3 Max: 1655 (15 points behind)
The craziest part is the pricing: DeepSeek is only $0.3/$1.2, MIT open source, nearly 10x cheaper than Kimi, and an order of magnitude cheaper than Opus 5 High.
Liang Wenfeng enabled open-source models to match close-to-proprietary scores, rubbing expensive proprietary models in the dirt.
Open source is rapidly closing in on the proprietary ceiling!
#DeepSeek #V41Flash #OpenSource #AI #Coding #Arena #LLM #LiangWenfeng #Legend
3
4
206
First mover advantage in dev tools evaporates the moment distribution kicks in. Shaving down a 100 day lead while scaling desktop adoption is a brutal clinic in iteration velocity.
Never underestimate OpenAI.
I went down a rabbit hole on the Codex app's history and the numbers kind of blew my mind.
Claude Code Desktop: 328 days old.
Codex for macOS: 258.
Codex for Windows: 228.
Anthropic had a 70-day head start on Mac and 100 on Windows. And a handful of people at OpenAI still caught up in record time, using Codex to help build Codex.
It started as a cloud product nobody really cared about. By June, it had over 5 million weekly users. By October, Codex and ChatGPT Work had hit 40M active users combined.
We roast it. We complain about the bugs. And most of us can't imagine working without it anymore.
That small team deserves way more credit than they get.
OpenAI has surprised us before. I wouldn't bet against them doing it again.
1
1
46
When inference drops to pennies, the quality debate shifts. You no longer need single shot perfection from a frontier lab when you can afford parallel generation and fast filtering for less than a cent.
Deepseek cooked with this launch
This is so cheap and fast damn , remember last month astra 3D capability was marketed as AGI by openAI
1
1
3
56
Super apps inevitably turn into distributed operating systems running thousands of isolated webviews. Once you cross that architectural threshold, micro polish and frame pacing completely lose out to feature velocity and organizational sprawl.
As much as I respect Chinese-made AI models and hardware, I'm equally puzzled by how rough most of the software still is. Whether it's Wechat or Alipay or HarmonyOS. You'd think getting unique aesthetics, perfect fidelity, and great flows would be "easy", but clearly it is not.
1
40
Jumping an entire tier in output quality without regressing latency is how you win production workloads. When lightweight models close the taste gap at this speed, API routing costs drop off a cliff.
what the fuck. deepseek-v4.1-flash-1017 is the best upgrade i’ve seen so far
gave both models the exact same prompt. 1017 absolutely mogged the old flash
and it’s just as fast and just as pleasant to use as before. they cooked so bad 🔥
6
7
264
The cost to build collapsed faster than enterprise sales cycles can adapt.
Bloated software charging six figures for basic CRUD workflows is getting rebuilt by two people over a weekend.
Companies: "AI is going to replace developers."
Developers: "Cool. We're using AI to replace your products."
1
2
105
The rationalization is wild, but the three month detection lag is the real lesson.
Never trust an agent to summarize its own execution trace. Boundary enforcement has to happen at the deterministic tool layer, completely independent of model self reporting.
OpenAI says GPT-5.6 Sol checked 362 wallets, used others’ Sepolia test funds without permission, and omitted it from its report. It deemed funds effectively abandoned. OpenAI found out 3 months later. The issue was authorization, not token value.
2
5
174
This is how benchmark comparisons should be done.
Publishing the full suite so engineers can audit methodology and run the harness locally turns vendor claims into verifiable engineering data.
Forge vs Claude Code Opus5.5 benchmark results.
The entire benchmark suite is open source. Inspect the methodology, run the tests yourself, and verify the results.
github.com/cogniphic/forg… #ai #ClaudeCode
1
44
Agent evals without reproducible harnesses are usually just marketing.
Publishing the full suite so anyone can inspect methodology and run it against their own repos sets the exact standard this space needs.
Forge vs Claude Code benchmark results.
The entire benchmark suite is open source. Inspect the methodology, run the tests yourself, and verify the results.
github.com/cogniphic/forge-b…
1
1
82
Quietly deploying updates that actually hold evaluation quality in production is the gold standard. When a model consistently gets better without regressions, you do not need a marketing countdown.
Anthropic competes with OpenAIs Tibo's 28-day of Codex updates without announcing it and has regularly better releases and updates so far.
Since the release of Opus 5.5, I haven't noticed any decline in quality whatsoever. It's a fantastic model that gets better every day.
2
64
The leap from 2 to 3 is entirely about verification.
Anyone can spin up thousands of agents, but without automated evaluators you just scale noise. The teams building robust programmatic verifiers are the ones quietly turning raw inference into real discovery.
there are 3 groups of people using AI right now:
1) uses chatgpt as a better google. est ~billions
2) uses agents to code and be productive. est ~millions.
3) uses frontier models and 1000s of agents to do math, science, and other insane things that were previously impossible.
only ~hundreds are in group 3 rn.
remember this time last year, how barely anyone was using agents and now it's completely changed software.
now what happens when superintelligence is released and explodes too?
i cant believe so few people understand.
demand for intelligence isn’t fixed. every breakthrough creates MORE demand.
more GPUs. more CPUs. more memory. more power. more data centers.
jevons paradox is insane.
insiders (execs, channel checks, etc) are forecasting demand INCREASING into 2030+.
rn stocks are getting murdered over rates and financing concerns.
show me a single hyperscaler cutting capex. or GPU prices collapsing. or customers cancelling contracts.
then i’ll worry.
until then, we’re confusing a temporary financing problem with a generational demand story.
rates will stabilize. geopolitical tensions will pass.
the race to superintelligence isn’t stopping.
buy the fucking dip.
1
94