activescott's Notes

Because vision, audio, and text inputs share the exact same weights, you no longer need to co-tune separate frozen encoders. Downstream adapter fine-tuning (such as LoRA) or full fine-tuning naturally updates the entire multimodal token loop in a single pass.

While many AI open source model providers are pursuing larger and more powerful models, Google is still giving attention to the smaller, more local side of the market. Today, the tech giant released Gemma 4 12B, an 11.95-billion-parameter open-weights model with permissive Apache 2.0 license optimized to execute locally on a standard enterprise laptop using just 16GB of VRAM or unified memory.

Traditional multimodal systems typically utilize discrete, separate encoders to translate audio waveforms and visual data into representations that the core language model can process.

This conventional approach inherently increases both inference latency and total memory consumption.

Gemma 4 12B radically alters this pipeline by functioning entirely without these secondary encoders. Instead, visual patches and raw audio waveforms are projected directly into the core large language model's embedding space through lightweight linear layers.

The vision encoder is replaced by a 35-million-parameter module utilizing a single matrix multiplication, while the audio encoder is eliminated entirely.

For enterprise engineering teams, this unified architecture delivers distinct operational advantages: lower latency for multimodal tasks, reduced VRAM requirements (down to 16GB — typical for laptops), and the ability to fine-tune the entire multimodal system in a single, cohesive pass.

Google has ensured that Gemma 4 12B is not an isolated experiment; it is ready for production. Weights are available on Hugging Face and Kaggle, and the model integrates seamlessly with industry-standard deployment frameworks such as vLLM, SGLang, MLX, and llama.cpp.

Interesting prompt for porting zig to rust in bun.

Meanwhile, Bun’s relationship with the Zig community has soured. Zig Software Foundation maintains a strict “no AI code” policy and previously rejected Bun’s attempts to upstream performance improvements as “technically unsound.” WinBuzzer reported on May 1 that “Anthropic-owned Bun is already paying the cost” of Zig’s LLM ban. The friction is accelerating the Rust exploration.

Zig was a deliberate choice in 2022. Its comptime evaluation and zero-runtime design minimized overhead when calling JavaScriptCore’s C++ interfaces. But Zig’s ecosystem remains narrow: no central package registry comparable to crates.io, a small hiring pool, and slow open-source contributor growth.

Rust is now the practical standard for systems programming. AWS Firecracker, Cloudflare’s Workers runtime (workerd), and Linux kernel drivers are built in Rust. The White House ONCD and NSA have formally recommended Rust for memory safety. As of 2026, choosing Rust means choosing a contributor pool orders of magnitude larger than Zig’s.

Bun’s Rust move is about ecosystem integration, not raw performance gains. More contributors, better security tooling, and more hireable engineers.

Docs for the OAK-D Lite.

Luxonis OAK-D Lite datasheet.

Kinda like twilio.

E.164 defines a general format for international telephone numbers. Plan-conforming telephone numbers are limited to only digits and to a maximum of fifteen digits. The specification divides the digit string into a country code of one to three digits, and the subscriber telephone number of a maximum of twelve digits.

As diplomatic efforts to strengthen the deal have stalled, three soldiers described to AP a sense of confusion in the embattled territory, with a lack of clarity on rules of engagement around the yellow line. Some commanders paid lip service to the agreement, the soldiers said, while privately voicing desire for the war in Gaza to continue. Sometimes, troops were too far away or acted too quickly to recognize who they were shooting, one soldier said — a concern echoed in comments from a whistleblower group of veterans.

When the ceasefire went into effect, Israel withdrew troops to a buffer zone demarcated by a yellow line, giving it control of just over half the strip. Under the agreement, Israeli forces are meant to complete a fuller withdrawal, though there’s no timeline for that. The U.S.-backed diplomat overseeing the truce says progress is deadlocked over the central sticking point of disarming Hamas, upon which all other issues — including Israeli withdrawals and reconstruction — hinge.

Since the ceasefire went into effect, more than 900 people have been killed in Gaza — dozens of those close to or over the yellow line, according to Gaza’s Health Ministry. The ministry doesn’t say how many are militants, but unarmed men and children have been among the dead.

soldiers who spoke to AP and Breaking the Silence — the whistleblower group that has collected troops’ testimonies throughout the war — say that at times soldiers were too far away, acting too quickly and under too much pressure to tell.

This week, Israeli Prime Minister Benjamin Netanyahu said Israel controls 60% of Gaza and the next step was to move to 70% control.

The soldiers told AP that on the ground, the ceasefire is elusive.

“We need to stop using this term,” one said of the word, ceasefire. “It’s not serving people that want to stop the war.”

Prefill processes all input tokens simultaneously. For a 4,096-token prompt, the attention computation involves large matrix multiplications across the full sequence length. This is compute-bound work. The GPU’s tensor cores are the bottleneck. On an H100 SXM, prefill achieves 200-400 arithmetic operations per byte of memory accessed. Utilization sits between 90% and 95%. The memory bandwidth, at 3.35 TB/s, is barely taxed.

Decode generates one token at a time. Each step reads the entire KV-cache from HBM to compute a single attention output. The tensor cores finish in microseconds and then wait for the next memory read. Arithmetic intensity drops to 60-80 ops/byte. GPU utilization falls to 20-40%. The tensor cores sit idle while the memory bus saturates.

Disaggregated inference runs prefill and decode on separate GPU pools connected by a fast network. A request arrives, gets routed to a prefill worker, which processes the full prompt and generates the KV-cache. That cache is then transferred over the network to a decode worker, which handles the autoregressive token generation.

Disaggregation is not free. The KV-cache produced during prefill has to move from the prefill GPU to the decode GPU over the network, and these caches are not small.

For a 70B parameter model using grouped-query attention with 80 layers, 8 KV heads per layer, 128 dimensions per head, stored in FP16: each token’s KV state is 327,680 bytes. A 4,096-token prompt produces 1.34 GB of KV-cache. That entire block has to transfer before the decode worker can begin generating.

Our tiered fee schedule rewards long-term commitment and exit fees decrease over time, reaching 0% after five years.

The FTSE Global Equity Shariah Index Series has been designed to be used as the basis of Shariah compliant investment products that meet the requirements of Islamic investors globally. Using the Large and Mid Cap stocks from the FTSE Global Equity Index Series as a base universe, constituents are then screened against a clear set of Shariah principles. Shariah screening is undertaken by Yasaar Limited, to create a Shariah compliant index series. The series has been fully certified as Shariah-compliant through the issue of a Fatwa (Islamic legal opinion) by Yasaar's principals. The FTSE USA Shariah Index contains USA stocks that meet these criteria. Their approach is described below.

In Trump’s second term, Michael Dell attended the president’s Invest America Roundtable event in June 2025. In December, Trump introduced the computer executive and his wife, Susan Dell, as they announced a $6.25 billion gift for 25 million U.S. children’s Trump Accounts. This month, Trump advised Americans to “go out and buy a Dell.”

On Wednesday, the company was awarded a $9.7 billion contract by the Pentagon to provide a suite of software to the U.S. military. The award came after what the Pentagon said was a competitive process, though the deal is being questioned by some experts in government.

The Pentagon contract, given to Dell Federal Systems, a part of the company that focuses on government, has raised eyebrows as a potential payback for the Trump Accounts donation.

“It looks terrible is the short answer,” said Greg Williams, director of the Center for Defense Information, part of the Project on Government Oversight. “Very publicly courting contributions to all sorts of projects of the president’s has created the very strong interest that they are in the business of soliciting contributions of various kinds in exchange either for access to the president or for outcomes that he may be able to influence.”